feat(C200): separar falas por pessoa nas gravações enviadas (Closes #1167)

This commit is contained in:
Francisco Solla 2026-09-19 12:05:35 -03:00
parent 78598906de
commit 7e2dd3a4cb
38 changed files with 19479 additions and 97 deletions

View file

@ -0,0 +1 @@
**C200 — Separar falas por pessoa nas gravações enviadas (2026-09-19):** as gravações do C199 saíam com a transcrição em bloco único e a assessoria não conseguia responder "o que fulano disse nesta gravação?". Entregou-se a camada de quem-falou-o-quê com agrupamento **acústico** e identificação **humana** (sem voiceprint, sem match voz↔pessoa, sem `Contact`): (1) **diarização** — o job ganhou uma etapa opcional atrás de um seam injetável (`SpeakerDiarizer`); o áudio extraído inteiro (uma passada, para a numeração ser global) vai ao provedor recomendado **AssemblyAI** `universal-2` + `speaker_labels` (US$0,17/h; chave `ASSEMBLYAI_API_KEY` ainda não configurada — sem chave/falha, a transcrição continua em bloco, nunca finge separação), e os turnos são alinhados aos segmentos do Whisper por sobreposição máxima (empate pelo ponto médio, sem sobreposição → turno mais próximo) com numeração `speaker-N` por primeira aparição; (2) **schema/migration** `20260919_032729_add_recording_speakers` — `recording_segment.speaker_key`, `recording.speaker_labels` (`{speakerKey,label}[]`), `recording.speaker_names` (`hasMany` derivado por hook, base da faceta) e `recording.speaker_labels_dropped`; o SQL duplicado de `reel.source_hash` que o snapshot defasado do C199 gerava foi removido à mão e o snapshot desta migration o corrige; (3) **reprocessamento não perde rótulo**: `reconcileSpeakerLabels` (puro) só mantém um rótulo em casamento mútuo 1:1 por sobreposição de tempo — split/merge descartam com aviso `speakerLabelsDropped` renderizado no próprio banner; (4) **faceta "Pessoa"** na fonte enviadas: param repetível `person` no contrato da lista (`contains` case-insensitive sobre `speaker_names`), chips removíveis (design cena 3), chip "aparece nesta gravação" no card e preservação do filtro ao submeter a busca; (5) **ação "Identificar falante"** (`POST .../gravacoes/[id]/falantes`) — rotula um agrupamento validando que a chave ainda existe na gravação, gate fail-closed do acervo (a escrita usa o predicado do dono; `advisor`/`leader` negados) e limpa o aviso de reconciliação; (6) **design C portado classe-a-classe** das cenas 1–5 (blocos por falante com avatar/estado/edição, diálogo de identificação, aviso âmbar de imprecisão sempre visível, mobile compacto); (7) **testes**: 17 unit novos (alinhamento/numeração/reconciliação/URL/where/provedor com fetch mockado), 4 int novos (faceta no join real, matriz de papéis da action, job com diarizador fake preservando rótulo e fallback com aviso) e 5 e2e novos na vertical; migração aplicada só em banco local, nenhuma gravação antiga reprocessada automaticamente.

View file

@ -0,0 +1,239 @@
# Impl: C200 — Separar falas por pessoa nas gravações enviadas
Status: aprovado
Atualizado em: 2026-09-19
Issue: #1167
Intenção: docs/plans/acervo-separacao-por-pessoa.md
Appetite restante: herdado (~2–3 dias eng; um outcome verificável — a assessoria identifica quem fala numa gravação enviada e filtra o acervo pelas falas de uma pessoa)
## Leitura da intenção
- **Outcome:** gravações novas do acervo (fonte "Gravações enviadas", C199) passam a ter a transcrição **dividida por agrupamento acústico** ("Falante 1", "Falante 2", …); a assessoria ouve e usa "Identificar falante" para nomear um agrupamento, valendo para todas as falas dele naquela gravação; a lista ganha a faceta **"Pessoa"** que recorta as gravações onde aquele rótulo aparece. A busca geral continua cobrindo a transcrição inteira.
- **O que NÃO negociar:** agrupamento é acústico e a identificação é **humana** (nada de voiceprint, match voz↔pessoa, inferência ou sugestão automática); rótulo é texto por gravação, **sem segundo cadastro de pessoa** (`Contact` intocado); rótulo não preenchido permanece "Falante N"; sem `Consent` novo (mídia interna de staff); sem score de confiança de "quem falou" na UI; gate fail-closed do acervo (`communicator` + `coordinator`/`candidate`; `advisor`/`leader` negados); nada de backfill automático das gravações antigas; sem UI nova de reprocessar/diarizar fora do design aprovado.
- **O que reavaliar** (hipóteses da intenção):
- "O import substitui os segmentos em bloco — um rótulo só sobrevive se o reprocessamento carregar/re-aplicar" (intenção): confirmado; a reconciliação é decisão desta impl (D6). Nota de realidade: hoje o lifecycle **não tem** caminho `ready → processing` (retry só `failed`, e falha não grava segmentos), então a reconciliação é um contrato do job testado no int, pronto para a ação futura de reprocessar/diarizar (§D8).
- "A fonte enviadas já tem o padrão de chips/omnibox" (achado do explorador): **não tem** — hoje é só `CampaignSearchInput`. A faceta nova reusa o **`CampaignHeaderFilterPopover`** (multi-seleção) com adaptação local, sem importar o omnibox da Câmara.
- "Reusar o pipeline de transcrição por chunks para diarizar": não serve — chaves por chunk não se ligam entre chunks (ver D2).
- "Provedor de diarização" (questão em aberto da intenção): resolvido em D1/D2 (provedor externo, áudio inteiro numa passada) e registrado como **GATE humano de custo/chave**.
## Abordagem recomendada
```mermaid
flowchart TD
U[Upload C199] --> J[recordingJob]
J -->|"ffmpeg: mp3 mono 16k / chunks 20min"| W["Deep Infra Whisper<br/>deepInfraTranscribeSegments"]
W --> M[mergeChunkTranscriptions]
J -->|"se há provedor: full.mp3 numa passada ffmpeg"| D["AssemblyAI universal-2<br/>language_code=pt, speaker_labels"]
D -->|"utterances speaker/start/end"| A["assignSpeakerKeys<br/>alinhamento por sobreposicao"]
M --> A
A --> T[["tx: recordingSegment.speakerKey<br/>+ recording.speakerLabels reconciliadas"]]
D -.->|"falha, timeout ou sem chave"| F["fallback honesto:<br/>transcricao sem agrupamento"]
F --> T
T --> R[(recording / recordingSegment)]
L["Lista ?source=enviadas&person=..."] -->|"speakerNames contains"| R
P["Detalhe /gravacoes/id"] -->|"blocos por falante"| R
DLG["Dialogo Identificar falante"] -->|"POST /gravacoes/id/falantes"| AC[labelRecordingSpeakerForActor]
AC --> R
```
**Opções consideradas:** A | B | C em cada decisão abaixo.
**Recomendação:** manter o Whisper/Deep Infra como dono da transcrição e adicionar uma **etapa opcional de diarização** sobre o áudio inteiro (AssemblyAI, `universal-2`, `speaker_labels`), **alinhando os turnos aos segmentos existentes** por sobreposição de tempo; chave do agrupamento no `recordingSegment`, rótulos na `recording` com a faceta denormalizada em `speakerNames` (`hasMany text` + `contains`, precedente `Speech.keywords`); reconciliação conservadora de rótulos no job + aviso; sem provedor/chave, a transcrição sai **sem agrupamento** (nunca fingir separação).
### Decisões de engenharia
**D1 — Provedor de diarização (pesquisa web, números oficiais).**
`Opções: A) AssemblyAI | B) Deepgram | C) Gladia | D) Speechmatics | E) diarização local no homeserver | F) LLM de áudio | G) Deep Infra (atual)`
`Recomendação: A — AssemblyAI é o único com preço/hora mais baixo entre os verificados, aceita arquivos de até 10 h (cobre plenária inteira numa passada), tem pt com WER ≤10% declarado, diarização explícita no modelo universal-2 e fluxo assíncrono com polling (submit → poll), que é exatamente a forma do job C199 (heartbeat por atualização do updatedAt).`
| Provedor | Diarização | Preço oficial (pré-gravado) | pt-BR | Limite por arquivo | Forma do fluxo |
| ---------------------------- | ------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------- | --------------------------------------------------------- | ------------------------------------------------------------------------------------------------- |
| **AssemblyAI** | `speaker_labels: true` (`universal-2`, `universal-3-5-pro`) | **US$0,15/h** (`universal-2`) + aditivo diarização standard **+US$0,02/h = US$0,17/h** ≈ US$0,0028/min | `pt` com "high accuracy (≤10% WER)"; `universal-3-5-pro` cita dialeto pt-BR | **10 h** por arquivo; 5 GB via URL/2,2 GB no `/v2/upload` | `POST /v2/upload` → `POST /v2/transcript` → poll; `utterances[]` com `speaker`/`start`/`end` (ms) |
| Deepgram | `diarize_model=latest` | Nova-3 pré-gravado **US$0,0043/min = US$0,258/h** (diarização incluída no pré-gravado) | `pt-BR` na lista do nova-3 | não verificado nesta sessão | `POST /v1/listen` único e síncrono (ou callback) |
| Gladia | incluída (pyannoteAI Precision-2) | async **US$0,61/h** | 100+ idiomas | **135 min** / 1 000 MB por request | upload → `/v2/pre-recorded` → poll |
| Speechmatics | incluída | Melia 1 batch **a partir de US$0,129/h** (em _production preview_ oficial) | 55+ idiomas | batch, limites no portal | submit → poll |
| Deep Infra (atual) | **não diariza** — catálogo ASR verificado ao vivo (6 modelos, nenhum com diarização) | US$0,00045/min no Whisper large-v3 | pt | chunks de 20 min (C199) | já é o ASR do job |
| Local (pyannote/sherpa-onnx) | sim (montagem própria) | sem API; custo de ops/CPU/modelos | depende do modelo | — | processo/serviço Python no homeserver |
| LLM de áudio (Gemini/OpenAI) | **não é diarização acústica** | por token | sim | — | áudio inteiro ao LLM |
- **A (recomendada):** `utilities/ai/assemblyAiDiarize.ts` como dono do provedor (forma de `deepInfraTranscribe.ts`), `ASSEMBLYAI_API_KEY` nova. A diarização roda só com chave; a política de delete do transcript remoto (`DELETE /v2/transcript/{id}`) é best-effort pós-leitura (higiene), e **jamais** se liga `speech_understanding.speaker_identification` (nome automático — proibido pelo guardrail).
- **Rejeitadas:**
- **B** porque custa ~50% mais por hora (US$0,258 vs US$0,17) e é chamada síncrona de horas — sem o polling nativo que dá heartbeat barato ao job; qualidade comparável e pt-BR ok, então fica como **alternativa de segunda escolha no gate**.
- **C** porque o async aceita no máximo **135 min/1 000 MB** por request (docs oficiais): uma plenária de horas exigiria fatiar e a numeração global entre fatias volta a ser um palpite (o problema que D2 elimina); além de ~3,6× o custo.
- **D** porque o modelo de menor preço (Melia 1) está em _production preview_ oficial e o tier de maior acurácia (Enhanced) não publica um rate card único e estável nas fontes oficiais consultadas; sem vantagem decisiva sobre A.
- **E** porque o homeserver não tem GPU e montar pyannote/sherpa-onnx adiciona runtime Python, download de pesos e um domínio de falha de ops sem seam de teste; revisitar só com custo/volume material ou mudança de política de dados.
- **F** porque LLM de áudio atribui falante probabilisticamente (alucina atribuição e nome), não entrega turnos com timestamp confiável e colidiria com a regra OPS123 (provider `openai/*` reservado a design no Teqo).
- **G** porque capacidades verificadas: nenhum dos 6 modelos ASR do catálogo Deep Infra diariza; uma chave nova no mesmo provedor não resolve.
**D2 — Escopo da diarização: áudio inteiro vs por chunk de 20 min.**
`Opções: A) uma passada no áudio inteiro | B) por chunk de 20 min com numeração global por ordem de primeira aparição | C) por chunk com re-associação por embedding entre chunks`
`Recomendação: A — o agrupamento é uma propriedade do áudio inteiro; diarizar por chunk gera chaves que não se ligam entre chunks e repetir "Falante 1" em cada chunk refere-se a pessoas diferentes (enganoso, contra o guardrail). O áudio extraído (mono 16 kHz, 32 kbps ≈ 14 MB/h) fica muito abaixo do limite de tamanho do provedor e dentro das 10 h de duração; acima de 10 h o job pula a diarização (fallback honesto).`
- **Rejeitadas:** **B** porque "Falante N" por chunk com numeração global por ordem de aparição **não** garante a mesma pessoa (sem biometria não há re-associação) e mente para a assessoria; **C** porque re-associação por embedding é voiceprint — proibida pelo produto e pela intenção.
**D3 — Papel do provedor: diarizar e alinhar vs substituir o ASR.**
`Opções: A) provedor só fornece turnos; alinhamento aos segmentos do Whisper | B) provedor passa a ser o ASR da gravação (utterances viram segmentos) | C) provedor só quando há chave; Whisper quando não há`
`Recomendação: A — o Whisper continua dono do texto, do searchText e dos timestamps por segmento (C199 pinado); o provedor entra apenas como camada de agrupamento. Falha do provedor não afeta a transcrição, e o alinhamento vira uma função pura testável sem rede.`
- **Rejeitadas:** **B** porque troca o dono do texto, muda o custo do caminho feliz (~10× por minuto), reescreve a busca e faria a falha do provedor derrubar a transcrição inteira; **C** porque duas qualidades de transcrição no mesmo acervo sem o usuário saber, e dois donos do texto quebrando pinos C199.
**D4 — Modelo de dados dos rótulos.**
`Opções: A) estado na recording (array `speakerLabels`) + `speakerNames` denormalizado (hasMany text) + chave no segmento | B) rótulo copiado em cada segmento | C) collection nova de rótulos | D) campo JSON`
`Recomendação: A — exatamente o que a tarefa fixou: recordingSegment.speakerKey (chave do agrupamento) + estado de rótulos na recording; a faceta "Pessoa" usa speakerNames (text hasMany → text[] no Postgres; precedente Activity.tags/Speech.keywords) com contains (ILIKE), funcionando no Payload 3.82.`
- A: `speakerLabels: { speakerKey, label }[]` (readOnly no admin; uma linha por agrupamento identificado) e hook `deriveSpeakerNames` preenchendo `speakerNames` (dedupe case-insensitive, ordem de aparição) — espelho de `RecordingSegment.deriveSearchText`. `speakerLabelsDropped: boolean` guarda o aviso de reconciliação (D6). Segmentos guardam só a chave.
- **Rejeitadas:** **B** porque renomear um agrupamento tocaria N segmentos e a faceta viraria um join caro; **C** é collection nova para um estado 1:N simples (custo de access/index/migration sem ganho) e foge da direção da tarefa; **D** não tem validação/índice nem precedente de query de faceta.
**D5 — Semântica da faceta "Pessoa".**
`Opções: A) `{ speakerNames: { contains: valor } }`OR entre valores selecionados | B)`{ speakerNames: { in: valores } }` exato | C) normalizar rótulo em slug na gravação`
`Recomendação: A — o precedente da casa (speechListFilters.ts:70, Speech.keywords contains) é case-insensitive e tolera variação ("Solla" casa "Dep. Jorge Solla"); a URL guarda o texto literal, a curadoria varia o rótulo e o recorte não fragmenta.`
- **Rejeitadas:** **B** porque a variante livre ("Solla" vs "Dep. Jorge Solla") fragmentaria a faceta (anti-goal de curadoria acidental); **C** porque cria uma segunda representação do rótulo e um merge implícito que a intenção cortou.
**D6 — Preservação de rótulos ao reprocessar.**
`Opções: A) reconciliação por sobreposição de tempo, conservadora 1:1, com flag + aviso | B) casar por igualdade de speakerKey | C) aceitar a perda silenciosa | D) bloquear o reprocessamento quando há rótulo`
`Recomendação: A — sem biometria não há re-associação entre execuções; carregar um rótulo só quando um agrupamento novo casa 1:1 por tempo com um antigo é o contrato honesto. Quem não casa fica sem nome e o job liga speakerLabelsDropped, que rende o aviso no detalhe.`
- A: o job lê labels + segmentos antigos **antes** de apagar, computa `reconcileSpeakerLabels` (puro: sobreposição acumulada por par, 1:1, ordenado por primeira aparição) dentro da mesma transação e grava `speakerLabels` reconciliadas + `speakerLabelsDropped`. O aviso vive **dentro** do banner âmbar da Cena 4 (uma linha condicional — ver §UI).
- **Rejeitadas:** **B** porque a chave é por execução e a mesma chave pode referir outra pessoa (o aceite manda re-vincular com aviso, não fingir); **C** viola o aceite; **D** porque o reprocessamento de hoje só existe para `failed` (que não tem rótulo) e, na ação futura, bloquear sem alternativa é pior que reconciliar e avisar.
**D7 — Fallback sem provedor/chave/timeout.**
`Opções: A) transcrição sem agrupamento (estado C199, sem "Falante") | B) falhar a gravação | C) criar um "Falante 1" único`
`Recomendação: A — o acervo continua útil e a ausência de blocos não afirma nada; é exatamente o detalhe C199 já aprovado, sem markup novo.`
- **Rejeitadas:** **B** porque faz a transcrição (que não depende do provedor) depender de um custo humano ainda não aprovado; **C** porque sugere um falante único onde há vários — fingir separação, proibido.
**D8 — Reprocessar/diarizar gravações antigas.**
`Opções: A) ação explícita por gravação agora | B) backfill automático | C) sem ação agora; follow-up com gatilho | D) UI de reprocessar junto com este item`
`Recomendação: C — a tarefa veta UI nova de reprocessar fora do design e a intenção veta backfill. A reconciliação (D6) fica pronta no job para quando a ação existir.`
- **Gatilho do follow-up:** chave do provedor ativa em produção **e** (pedido operacional de diarizar o acervo antigo **ou** passarem ~30 gravações novas sem diarização). Aí o designer estende o artefato **antes** do markup (trigger a). **Débito absorvido pelo D8 (triagem do /simplify):** quando a ação de reprocessar existir, a semântica de limpar `speakerLabelsDropped` no rótulo precisa considerar rótulos ainda perdidos (hoje inalcançável: não há `ready → processing`).
- **Débito absorvido pelo C202:** a action de rótulo gateia no action-level com `canReadCommunicationCatalog` (a collection já exige `canUploadRecording`); a troca pelo predicado de escrita do dono foi adicionada ao plano `docs/plans/gates-gravacoes-predicados.md` (#1186).
- **Rejeitadas:** **B** porque o custo e o risco de mexer em rótulos não justificam varrer tudo; **A/D** porque contrariam a restrição de superfície do item.
## Componentes / mudanças
- **`src/lib/recordingDiarization.ts`** (novo, puro/client-safe): `DiarizedTurn`, `assignSpeakerKeys` (sobreposição máxima; empate pelo ponto médio; sem sobreposição → turno mais próximo; numeração global `speaker-1…` por primeira aparição), `reconcileSpeakerLabels` (D6), `speakerNamesFromLabels`, `recordingSpeakerDefaultLabel(n)` → `Falante ${n}`, `RECORDING_DIARIZATION_MAX_SECONDS = 10h` (limite do provedor citado em comentário). Mesmo padrão de dono dos puros de `lib/recordingTranscription.ts`.
- **`src/lib/recordingTranscription.ts`** (editar): `buildRecordingFullAudioFfmpegArgs` (mp3 mono 16k/32k sem segmentação) ao lado do args de chunks; merge/searchText intocados.
- **`src/lib/recording.ts`** (editar): `RECORDING_SPEAKER_LABEL_MAX_LENGTH = 120`, mensagens pt-BR do diálogo/action, nada de status novo.
- **`src/utilities/ai/assemblyAiDiarize.ts`** (novo, `server-only`): dono do provedor (molde `deepInfraTranscribe.ts:98-137`) — upload → transcript (`language_code: 'pt'`, `speaker_labels: true`, sem identificação automática) → poll com deadline de 60 min e intervalo de 5 s; normalização ms→s; nunca lança (`{ ok, turns }` / `{ ok:false, error, status }`); `DELETE` best-effort do transcript; chave ausente → 503. Exporta `assemblyAiDiarize` e `configuredSpeakerDiarizer()` (null sem chave).
- **`src/utilities/recordings/recordingJob.ts`** (editar): `runRecordingJob(payload, id, transcribe = deepInfraTranscribeSegments, diarize = configuredSpeakerDiarizer())`; extrai `full.mp3` só quando `diarize` existe e a duração ≤ 10 h, chama o provedor (heartbeat `step:'transcribing'`), alinha com `assignSpeakerKeys`, e na transação já existente (`recordingJob.ts:191-227`) lê o estado antigo, apaga/recria segmentos com `speakerKey` e grava `speakerLabels`/`speakerLabelsDropped`. Falha/timeout do provedor **não** marca `failed`.
- **`src/collections/Recording.ts`** (editar): `speakerLabels` (array `{ speakerKey, label }`, readOnly), `speakerNames` (`text` `hasMany`, readOnly), `speakerLabelsDropped` (`checkbox`, readOnly) + hook `deriveSpeakerNames` (fallback `originalDoc` como `deriveSearchText`).
- **`src/collections/RecordingSegment.ts`** (editar): `speakerKey` (`text`, opcional, readOnly).
- **`src/utilities/recordings/recordingListUrl.ts`** (editar): param `person` (repetível) no `RecordingListState`, parse com `allParamValues` + trim + cap 120 + dedupe, serialize com `append`, `toggleRecordingPerson`/`removeRecordingPerson` puros (page reset) — o "adapter" da faceta mora aqui em vez de um módulo espelho de `speechOmnibox.ts` (depth check).
- **`src/utilities/recordings/recordingListFilters.ts`** (editar): `buildRecordingListWhere` AND-e `searchText like` + `{ or: persons.map(p => ({ speakerNames: { contains: p } })) }`.
- **`src/utilities/recordings/recordingViewModels.ts`** (editar): `RecordingSpeakerGroupViewModel` (`key`, `index`, `defaultLabel`, `label`, `isIdentified`, `segments`), `RecordingDetailViewModel.speakerGroups` + `speakerLabelsDropped`; `RecordingListItemViewModel.matchedPersons`; agrupamento por primeira aparição; caminho sem chave renderiza a lista plana C199.
- **`src/utilities/recordings/recordingPageData.ts`** (editar): selects (`speakerNames`, `speakerLabels`, `speakerLabelsDropped`, `speakerKey`), `loadRecordingFilterOptions` (select de `speakerNames` com `pagination:false`, dedupe e sort pt-BR — molde `speechPageData.ts:85-121`), `matchedPersons` por linha (puro `contains` case-insensitive espelhando o where).
- **`src/app/(campaign)/campanha/actions/recording.ts`** (editar): `labelRecordingSpeakerForActor({ recordingId, speakerKey, label })` — gate pelo **predicado de escrita do dono** (C202, §Dependências), valida que a chave existe em algum segmento da gravação, faz read-modify-write do array `speakerLabels` e limpa `speakerLabelsDropped` (o ato humano reconhece/revisa); 1 collection → sem transação.
- **`src/lib/schemas/recording.ts`** (editar): `recordingSpeakerLabelRequestSchema` (id, `speakerKey` `^speaker-\d{1,3}$`, label 1..120 trim) + mensagens seguras.
- **Rotas** (`.../acervo/gravacoes/[id]/falantes/route.ts`, novo): `POST` via `campaignJsonMutationRoute` (sem allowlist de conventions — rota wrapper, não crua), molde de `[id]/retry/route.ts`; `types.ts` ganha `RecordingSpeakerLabelResponse`; `lib/campaignPaths.ts` ganha `campaignRecordingSpeakersHref`.
- **UI** (`src/components/campaign/recording/`): `RecordingSpeakerTranscript.tsx` (novo, client; blocos + banner âmbar + botões), `RecordingSpeakerDialog.tsx` (novo, client; molde `SpeechCutDeleteDialog` + `postCampaignJson`), `RecordingPersonFilter.tsx` (novo, client; `CampaignHeaderFilterPopover` + `useCampaignListFilterNavigation`), edições em `RecordingDetailPlayer.tsx` (renderiza grupos quando existem; senão o `<ol>` atual), `RecordingResultCard.tsx` (chip de pessoa) e `acervo/page.tsx` (hidden inputs de `person` no form de busca, linha da faceta, vazio com filtro).
- **Manifest e2e** (`scripts/lib/e2e-affected-manifest.mjs`): **sem mudanças esperadas** — `src/utilities/ai` (linhas 261-265) já mapeia `campaignSpeechAcervo`, e `src/utilities/recordings`/`src/components/campaign/recording`/`src/lib/recording` (322-336) já cobrem os arquivos novos; confirmar no gate.
- **Migration:** `add_recording_speakers` (abaixo).
- **Access / Consent:** sem `Consent` novo; `utilities/access/recordings.ts` intocado neste item (C202 é o dono da ampliação de gate); a action de rótulo é a única superfície nova e usa o predicado de escrita do dono.
- **UI:** Impeccable **C** — port do artefato + shells (`CampaignHeaderFilterPopover`, `CampaignListPendingBoundary`, `CampaignListEmptyState`, `CampaignListFooter`, `Dialog`, `Alert`, `Spinner`); shape→craft→critique→polish no ciclo.
### Dados → forma (se aplicável)
- **N/A por decisão da intenção** (`acervo-separacao-por-pessoa.md:59`): filtro por pessoa é recorte de conteúdo, não métrica/agregado; nenhuma confiança numérica é exposta (o produto proíbe score de "quem falou"). Nada a apresentar; nenhuma forma nova.
## UI e port do design (fonte de verdade: `acervo-separacao-por-pessoa-ui-design.html`)
**Coberto pelo artefato — port classe-a-classe:**
1. **Cena 1 (detalhe desktop, agrupado):** grid `grid-cols-[minmax(0,1.05fr)_minmax(430px,.95fr)]`, mídia à esquerda (player + `Baixar` + banner informativo **literal** "Os falantes são agrupados automaticamente pelo áudio. A identificação de cada pessoa é feita pela equipe.") e coluna direita com heading uppercase "Transcrição por falante", hint "Clique para posicionar" e `article.speaker-block` por grupo: head (`size-7` avatar, `Falante N`/rótulo, sub "Agrupamento sem identificação"/"Identificação feita pela equipe", botão outline "Identificar falante"/"Editar identificação") e linhas `seg` (grid `3.25rem 1fr`, `data-start-seconds`, ativo com `bg-muted`, hora em `text-primary` quando ativo, texto com `SpeechHighlightParts`).
2. **Cena 2 (diálogo):** `Dialog` com título "Identificar falante", descrição `Dê um nome ao agrupamento “Falante N”.` (sempre o rótulo default), caixa informativa "Ouça um trecho e informe quem é a pessoa. O acervo não sugere nem identifica pessoas automaticamente.", label "Nome do falante", placeholder "Digite o nome do falante", helper "O nome será aplicado a todas as falas deste agrupamento nesta gravação.", botões "Cancelar"/"Salvar identificação"; erro inline (molde retry/delete), `router.refresh()` no sucesso.
3. **Cena 3 (lista com faceta):** mantém `RecordingsSearchForm` (`CampaignSearchInput`, C199) **com hidden inputs dos `person` ativos** (submeter a busca não pode perder o filtro) e adiciona a linha de chips (botão outline "Pessoa: {rótulo}" com X) via `CampaignHeaderFilterPopover` multi-seleção; card com o chip `{rótulo}` + "aparece nesta gravação" sob o excerpt.
4. **Cena 4 (aviso de imprecisão):** banner âmbar com título "A separação por falante pode estar imprecisa" e corpo "O agrupamento é automático e aproximado. Ruído e falas sobrepostas podem colocar trechos no agrupamento errado. A equipe continua responsável por identificar cada falante." — **sempre visível no detalhe agrupado** (sem score de qualidade, D1/anti-goal; o artefato não define gatilho). A linha condicional de reconciliação (D6) entra **dentro desse mesmo banner** quando `speakerLabelsDropped`: "Esta gravação foi reprocessada e algumas identificações não puderam ser mantidas. Confira os falantes." — uma frase, sem componente/superfície nova.
5. **Cena 5 (mobile 390 px):** mesmo detalhe em coluna única, head compacto do bloco (avatar + nome + botão "Identificar"/ícone de edição com `aria-label` "Editar identificação"), banner informativo curto **literal** "Agrupamento automático e aproximado. A identificação é feita pela equipe." e hint "Toque para posicionar".
**Decisões de port fora do literal (baratas, revisáveis):**
- **Subtítulo da página na Cena 3** ("Gravações enviadas"): mantém o subtítulo aprovado do C199 ("Busque nas falas da Câmara ou nas gravações da equipe.") + alternador; a Cena 3 é shorthand do estado da fonte e mudar o chrome regride o pin do C199 sem ganho de produto.
- **Vazio com filtro de pessoa**: reusa `CampaignListEmptyState` com a copy da casa ("Tente outro termo, remova filtros ou limpe a busca.") — estado não desenhado, shell aprovada.
- **Fallback sem agrupamento**: renderiza o detalhe C199 **como está** (sem copy nova de "não diarizado"). A tarefa manda disparar o designer se precisarmos de superfície nova — **não precisamos**; se surgir demanda de explicar a ausência, é trigger a (designer antes do markup), registrado como follow-up (§D8).
- Dialog sobre grupo já identificado reusa a copy da Cena 2 com o rótulo default; o trigger é que muda ("Editar identificação").
- `role="tablist"` inexistente aqui; o alternador C199 fica intocado.
## Migration
- Nome: `pnpm migrate:create add_recording_speakers` → gera coluna `recording_segment.speaker_key` (`text`), tabela `recording_speaker_labels` (`id`, `_order`, `_parent_id` FK `recording` cascade, `speaker_key`, `label`), coluna `recording.speaker_names` (`text[]`) e `recording.speaker_labels_dropped` (`boolean`). **Revisar o SQL**; sem índices novos (a faceta é `contains` sobre `text[]` em tabela pequena e o `Speech.keywords` não tem índice dedicado). Referência de forma: `20260715_163458_initial` (arrays) e `20260918_201009_add_reel`.
- **HARD-STOP de aprovação humana explícita, mesmo em sessão `--auto`, e local-only:** criar/aplicar a migração apenas no banco local do worktree (`pnpm migrate`), **nunca** tocar banco remoto/prod e **nunca** editar migração antiga. Parar, apresentar o SQL gerado e aguardar aprovação antes de seguir para o job/UI.
## Invariantes
- Local API em contexto de usuário sempre com `overrideAccess: false` + `user` (lista, detalhe, action de rótulo); bypass admin só no pipeline que é dono do row, com comentário (forma `recordingJob.ts:54-56`).
- Escrita multi-collection em transação com `req: { transactionID }` (`withPayloadTransaction`): substituição de segmentos + estado de rótulos + `searchText`/status na mesma transação; a action de rótulo é 1 collection.
- Sem `Consent` novo; sem `Contact`; rótulo é texto; nenhuma URL pública; `/arquivo` e matriz de papéis intocados.
- Gate fail-closed do acervo; a escrita nova usa o predicado de escrita do dono (C202) — nunca `canReadCommunicationCatalog` gateando escrita; `advisor`/`leader` negados (matriz no int/e2e).
- Sem biometria/voiceprint/unificação entre gravações; sem score/confiança na UI; sem inferência de identidade.
- "Edit the owner, don't twin": puros no `lib` dono da pipeline, provedor no `utilities/ai` dono, where/estado no `recordings/`, sem módulo top-level novo em `src/utilities/`.
- Identificadores em inglês; copy/labels pt-BR; valores de URL em pt-BR só os existentes (`source=enviadas`, `falantes`).
- `pnpm generate:types` após o schema; `pnpm gate:fast` durante; `pnpm push` como entrega; **este plano e `docs/changelog/2026-09-19-c200.md` entram no commit da entrega**.
## Fases verificáveis
1. **Puros + schema + migration (HARD-STOP)** (~1 dia).
- `lib/recordingDiarization.ts`, `lib/recordingTranscription.ts` (+full), `recordings/recordingListUrl|Filters|ViewModels|PageData`, `lib/recording.ts`, collections + hook; `pnpm generate:types`.
- Unit: alinhamento (sobreposição, empate, sem sobreposição→mais próximo, vazio), numeração global por primeira aparição, reconciliação (1:1 carrega; split/merge/orfão dropa + flag), `speakerNames` derivado, guard de 10 h, ffmpeg args do full, parse/serialize `person` (multi/dedupe/cap), where `contains`+q AND, `matchedPersons`.
- Int: matriz de papéis e persistência das três colunas novas; **validação da query da faceta `contains` em `text[]` no Payload 3.82** (precedente `Speech.keywords` e `Activity.tags`).
- **Migration `add_recording_speakers` — HARD-STOP humano, local-only.**
2. **Provedor + job + ação** (~1 dia).
- `assemblyAiDiarize` (seam injetável na assinatura do job), job com reconciliação na transação, action/rota/schema/paths.
- Int: job com diarizador fake (feliz multi-chunk → `speakerKey` por segmento + labels preservadas; falha do provedor → `ready` sem agrupamento; reconciliação com fixtures antigas; chave inexistente → fallback), `labelRecordingSpeakerForActor` (papéis; chave desconhecida; label vazio/longo; limpa `speakerLabelsDropped`), sem chamada real de rede.
3. **UI** (~1 dia).
- Cenas 1–5 + hidden inputs + faceta + cards; ação de rótulo no diálogo.
- Unit: view models dos grupos (ordem, `Falante N`, rótulo aplicado, flag), cena/copy constants; e2e cobre o markup.
4. **Gates + GATE humano** (~0.5 dia).
- Estender `tests/e2e/campaignSpeechAcervo.e2e.spec.ts` (describe C199): fixture `ready` com segmentos com `speakerKey` + `speakerLabels`; detalhe renderiza blocos/"Falante N"/rótulo/aviso; lista `?person=` recorta e o card mostra o chip; fallback sem `speakerKey` mantém `data-start-seconds` do C199; advisor negado na rota nova.
- `pnpm gate:fast`; `pnpm test:e2e:affected`; manifest/conventions (confirmar sem mudança); **smoke manual do provedor com 1 gravação real** (chave aprovada); changelog; push.
Quota: ~3–3,5 dias eng, dentro do appetite herdado; se apertar, o corte é o polish do port (a estrutura das cenas 1/2/3 fica antes do acabamento).
## Rabbit holes / Não escopo (engenharia)
- Voiceprint/embedding para re-associar falantes entre execuções/gravações; unificação automática de "mesma pessoa".
- Cadastro/diretório de pessoas ou vozes; qualquer uso de `Contact`.
- Edição/correção de transcrição, alinhamento por palavra, refino de sobreposição.
- Backfill de diarização do acervo antigo e UI de "Diarizar" (follow-up com gatilho, §D8).
- Multi-provedor/adapter genérico de diarização (uma interface, um provedor — sem volatilidade que justifique mais).
- Webhook/callback do provedor (o poll do job basta; callback criaria superfície pública nova).
- Score/confiança de falante na UI (proibido); detecção automática de "qualidade ruim".
- Fila/worker, tempo real, streaming, cortes por falante, busca semântica (C192) e merge de fontes.
## Riscos e mitigação
- **Custo/chave do provedor não aprovados (GATE).** Sem chave, tudo degrada para a transcrição sem agrupamento (D7); a feature nova só liga após a aprovação — nenhum teste depende de rede.
- **`speaker_labels` não suportado para `pt` na conta real.** Os docs oficiais da AssemblyAI listam `pt` com alta acurácia no `universal-2` e diarização nos modelos `universal-2`/`universal-3-5-pro`, mas a matriz feature×idioma não é pública: **smoke manual na fase 4** antes do rollout; falha honesta = fallback.
- **Qualidade do agrupamento em plenária ruidosa/sobreposta.** Banner da Cena 4 sempre visível + identificação humana + nenhum claim automático; a copy do provedor recomenda ≥30 s de fala por pessoa (citada nos docs).
- **Diarização > 10 h ou timeout de poll.** Guard de duração + deadline de 60 min com heartbeat; resultado = fallback sem agrupamento, nunca `failed`.
- **Alinhamento errado na fronteira de turnos.** Atribuição por sobreposição máxima com desempate pelo ponto médio; unidade cobre; aviso de imprecisão cobre o resto.
- **Reconciliação conservadora demais (rótulo perdido).** Só perde quando não há mapeamento 1:1; `speakerLabelsDropped` dispara o aviso; teste int garante carregar no caso trivial (mesmos tempos).
- **Corrida no read-modify-write de `speakerLabels`.** Aceita para ferramenta interna de equipe única (uma gravação editada por vez na prática); se doer, collection dedicada de rótulos é o follow-up.
- **Faceta cara conforme o catálogo cresce.** Select de `speakerNames` + dedupe é O(n) com n pequeno; gatilho: acervo > ~2 000 gravações ou opções > ~200 (pré-computar opções/`distinct`).
- **Regressão do C199.** O caminho sem `speakerKey` mantém o markup exato do `<ol>` atual; os testes C199 (unit/int/e2e) continuam a rede.
- **Vazamento de dados ao provedor externo.** Áudio de eventos públicos da equipe, sem PII de terceiros; delete best-effort do transcript remoto; política de retenção validada no gate.
## GATE humano — provedor, custo e chave (assumido; validar antes do rollout)
- **Assumido:** provedor **AssemblyAI** (`universal-2` + `speaker_labels`), custo **US$0,17/h** de áudio (US$0,0028/min) — fonte oficial: pricing AssemblyAI (Universal-2 US$0,15/h; diarização async standard +US$0,02/h) e docs de Speaker Diarization/limites (10 h, 5 GB; upload 2,2 GB).
- **Custo projetado:** 10 h/mês ≈ US$1,70; 50 h/mês ≈ US$8,50; 200 h/mês ≈ US$34,00.
- **Pendente de aprovação humana (mesmo com `--auto`):** criar conta/chave na AssemblyAI, definir `ASSEMBLYAI_API_KEY` no env de produção (e staging se desejado), aceitar o envio do áudio a terceiro e o orçamento acima; rodar o smoke real. **Nada disso bloqueia a implementação** (seam injetável + testes sem rede); bloqueia só o rollout.
- **Alternativa registrada no gate:** Deepgram (US$0,0043/min = US$0,258/h, diarização incluída, pt-BR) — se a aprovação preferir chamada síncrona única; a troca exige só um novo dono em `utilities/ai/` (a interface do seam não muda).
## Aceite de engenharia
- [ ] Aceite de produto da intenção coberto: transcrição dividida por "Falante N" nas gravações novas com provedor; "Identificar falante" vale para todo o agrupamento; faceta "Pessoa" recorta a fonte enviadas; busca geral intocada; rótulo vazio permanece "Falante N"; nunca sugere/infere identidade; reprocessar não perde rótulo sem re-vinculação avisada; antigas não são reprocessadas automaticamente.
- [ ] Invariantes AGENTS/engineering-standards: `overrideAccess:false` + `user`; transação multi-collection; sem `Consent`/`Contact`; `advisor`/`leader` fail-closed; sem biometria/score; sem módulo top-level novo em `utilities/`; migração nova sem editar antigas; plano + changelog no commit; `pnpm push`.
- [ ] Testes de domínio onde access/write paths mudam: puros da diarização/reconciliação/URL/where/view models no unit; job com seam injetado + action de rótulo com matriz de papéis + query da faceta no int; markup do detalhe/liste no e2e existente.
## Self-score de decision-quality
**4,5/5.** As decisões caras têm alternativas honestas e rejeições ancoradas: provedor com número oficial de quatro candidatos + limitações reais (135 min da Gladia, preview da Melia, 10 h/5 GB da AssemblyAI), escopo da diarização (áudio inteiro vs chunk — o modo chunk mente e foi rejeitado), papel do provedor (não substitui o ASR dono), modelo dos rótulos com precedente de query validado (`hasMany` + `contains` no Payload 3.82), semântica da faceta, reconciliação conservadora com aviso e fallback honesto. Cabe no appetite herdado (fases somam ~3–3,5 dias, tracer nos puros+schema cedo) e reusa shells/donos existentes (`CampaignHeaderFilterPopover`, `withPayloadTransaction`, `campaignJsonMutationRoute`, moldes de `deepInfraTranscribe`, `SpeechCutDeleteDialog`, `speechListFilters`). Perde 0,5 porque duas coisas ficam fora do código e viram parada dura/validação: a **migração** (HARD-STOP explícito, local-only) e o **custo/chave do provedor** (assumido; gate humano com smoke real), além de duas escolhas de port que dependem de ratificação no gate (banner da Cena 4 sempre visível sem score; subtítulo do C199 mantido na Cena 3).

View file

@ -18,6 +18,7 @@ Expor no dono (`utilities/access/recordings.ts`) um predicado de papel compartil
- `POST .../gravacoes/enviar` (`route.ts` — gate cru de sessão e origem).
- `retryRecordingForActor`, `deleteRecordingForActor` (`actions/recording.ts`).
- `labelRecordingSpeakerForActor` (`actions/recording.ts`; absorvido do C200 #1167 — a action gateia no action-level com `canReadCommunicationCatalog` enquanto a collection usa `canUploadRecording`).
- `getRecordingStatusesForActor` é leitura: mantém `canReadCommunicationCatalog`.
Rejeitadas: duplicar o corpo dos predicados nas rotas (gêmeo); criar um quarto predicado (o dono já tem os três).

View file

@ -0,0 +1,32 @@
import { NextResponse } from 'next/server'
import { labelRecordingSpeakerForActor } from '@/app/(campaign)/campanha/actions/recording'
import {
RECORDING_FORBIDDEN_MESSAGE,
RECORDING_GENERIC_ERROR_MESSAGE,
RECORDING_NOT_FOUND_MESSAGE,
RECORDING_SPEAKER_UNKNOWN_MESSAGE,
recordingSpeakerLabelRequestSchema,
} from '@/lib/schemas/recording'
import { campaignJsonMutationRoute } from '@/utilities/campaignJsonMutationRoute'
import type { RecordingSpeakerLabelResponse } from '../../types'
export const dynamic = 'force-dynamic'
/** C200 — the team identifies one speaker cluster of the recording (human). */
export const POST = campaignJsonMutationRoute(
{
bodySchema: recordingSpeakerLabelRequestSchema,
safeMessages: [
RECORDING_FORBIDDEN_MESSAGE,
RECORDING_NOT_FOUND_MESSAGE,
RECORDING_SPEAKER_UNKNOWN_MESSAGE,
],
genericMessage: RECORDING_GENERIC_ERROR_MESSAGE,
},
async (body) => {
const result = await labelRecordingSpeakerForActor(body)
return NextResponse.json<RecordingSpeakerLabelResponse>({ status: 'success', ...result })
},
)

View file

@ -116,6 +116,8 @@ export default async function RecordingDetailPage({
downloadHref={recording.downloadHref}
failureMessage={recording.failureMessage}
segments={recording.segments}
speakerGroups={recording.speakerGroups}
speakerLabelsDropped={recording.speakerLabelsDropped}
initialSeconds={initialSeconds}
/>

View file

@ -19,3 +19,8 @@ export type RecordingDeleteResponse =
export type RecordingRetryResponse =
| { status: 'success'; recording: RecordingViewModel }
| { status: 'error'; message: string }
/** Wire contract of `POST .../gravacoes/[id]/falantes` (C200) — label a cluster. */
export type RecordingSpeakerLabelResponse =
| { status: 'success'; labeled: true }
| { status: 'error'; message: string }

View file

@ -5,6 +5,7 @@ import { redirect } from 'next/navigation'
import { getPayload } from 'payload'
import { AcervoSourceToggle } from '@/components/campaign/recording/AcervoSourceToggle'
import { RecordingPersonFilter } from '@/components/campaign/recording/RecordingPersonFilter'
import { RecordingResultList } from '@/components/campaign/recording/RecordingResultList'
import { RecordingStatusRefresher } from '@/components/campaign/recording/RecordingStatusRefresher'
import { RecordingUploadDialog } from '@/components/campaign/recording/RecordingUploadDialog'
@ -77,9 +78,19 @@ const AcervoHeader = ({ source }: { source: 'camara' | 'enviadas' }) => (
</div>
)
const RecordingsSearchForm = ({ query }: { query?: string }) => (
const RecordingsSearchForm = ({
query,
people,
}: {
query?: string
people?: readonly string[]
}) => (
<form method="get" action={CAMPAIGN_COMMUNICATION_ACERVO} className="flex flex-col gap-1.5">
<input type="hidden" name="source" value="enviadas" />
{/* The "Pessoa" facet survives a new textual search. */}
{(people ?? []).map((person) => (
<input key={person} type="hidden" name="person" value={person} />
))}
{/* The shared input owns the accessible label; this is its visible twin. */}
<p className="text-sm font-medium">Buscar nas gravações enviadas</p>
<CampaignSearchInput
@ -101,10 +112,14 @@ const RecordingsSource = ({
data: Awaited<ReturnType<typeof loadRecordingsPageData>>
}) => {
const hasSearch = Boolean(data.state.q)
const hasPeople = (data.state.people?.length ?? 0) > 0
const hasFilters = hasSearch || hasPeople
return (
<CampaignListPendingBoundary>
<RecordingsSearchForm query={data.state.q} />
<RecordingsSearchForm query={data.state.q} people={data.state.people} />
<RecordingPersonFilter state={data.state} options={data.filterOptions.people} />
<RecordingStatusRefresher
recordings={data.rows.map((row) => ({ id: row.id, status: row.status }))}
@ -116,21 +131,27 @@ const RecordingsSource = ({
) : (
<CampaignListEmptyState
className="border-solid"
icon={hasSearch ? SearchIcon : VideoIcon}
icon={hasFilters ? SearchIcon : VideoIcon}
title={
hasSearch
? `Nenhuma gravação encontrada para "${data.state.q}"`
: 'Nenhuma gravação enviada ainda'
: hasPeople
? 'Nenhuma gravação encontrada com esse filtro'
: 'Nenhuma gravação enviada ainda'
}
description={
hasSearch
? 'Tente outro termo ou limpe a busca para ver todas as gravações enviadas.'
: 'Envie um arquivo de vídeo para acompanhar a transcrição e pesquisar o conteúdo no acervo.'
: hasPeople
? 'Tente outro termo, remova filtros ou limpe a busca.'
: 'Envie um arquivo de vídeo para acompanhar a transcrição e pesquisar o conteúdo no acervo.'
}
>
{hasSearch ? (
{hasFilters ? (
<Button asChild variant="outline" className="min-h-11">
<Link href={buildAcervoSourceHref('enviadas')}>Limpar busca</Link>
<Link href={buildAcervoSourceHref('enviadas')}>
{hasPeople ? 'Limpar filtros' : 'Limpar busca'}
</Link>
</Button>
) : (
<RecordingUploadDialog triggerLabel="Enviar gravação" />

View file

@ -12,14 +12,17 @@ import {
RECORDING_FORBIDDEN_MESSAGE,
RECORDING_NOT_FOUND_MESSAGE,
RECORDING_RETRY_NOT_FAILED_MESSAGE,
RECORDING_SPEAKER_UNKNOWN_MESSAGE,
recordingDeleteRequestSchema,
recordingRetryRequestSchema,
recordingSpeakerLabelRequestSchema,
recordingStatusRequestSchema,
type RecordingStatusRequest,
} from '@/lib/schemas/recording'
import type { CampaignUser } from '@/payload-types'
import { getCampaignActionContext } from '@/utilities/campaignActionContext'
import { onPayloadTransactionCommit, withPayloadTransaction } from '@/utilities/payloadTransaction'
import { acquireTextAdvisoryLocks } from '@/utilities/postgresTransactionLocks'
import { reapStaleRecording } from '@/utilities/recordings/recordingJob'
import { startRecordingJobInBackground } from '@/utilities/recordings/recordingScheduler'
@ -39,6 +42,7 @@ const recordingSelect = {
durationSeconds: true,
media: true,
updatedAt: true,
speakerLabels: true,
} as const
const loadRecordingForActor = async (
@ -138,6 +142,87 @@ export const getRecordingStatusesForActor = async (
return result.docs.map(toRecordingViewModel)
}
/**
* C200 — names one acoustic cluster of a recording. Human curation only: the
* key must belong to a segment of this recording (a stale dialog cannot invent
* a group), and the label is text on the recording — never a `Contact`. The
* write triggers the collection's `deriveSpeakerNames` hook, which is what the
* "Pessoa" facet filters on.
*
* Runs inside a transaction under an advisory lock of the recording and
* re-reads the labels after acquiring it: two curators labeling different
* clusters at the same time would otherwise each write back a stale array and
* silently lose one label. The lock closes that read-modify-write race.
*/
export const labelRecordingSpeakerForActor = async (input: {
recordingId: number
speakerKey: string
label: string
}): Promise<{ labeled: true }> => {
const parsed = recordingSpeakerLabelRequestSchema.parse(input)
const { payload, actor } = await getCampaignActionContext()
if (!canReadCommunicationCatalog(actor.role)) throw new Error(RECORDING_FORBIDDEN_MESSAGE)
const current = await loadRecordingForActor(payload, actor, parsed.recordingId)
if (!current) throw new Error(RECORDING_NOT_FOUND_MESSAGE)
await withPayloadTransaction(payload, async ({ req }) => {
await acquireTextAdvisoryLocks(payload, req, [`recording-speakers:${parsed.recordingId}`])
const fresh = await payload.find({
collection: 'recording',
where: { id: { equals: parsed.recordingId } },
depth: 0,
limit: 1,
pagination: false,
select: { speakerLabels: true },
req,
user: actor,
overrideAccess: false,
})
const recording = fresh.docs[0]
if (!recording) throw new Error(RECORDING_NOT_FOUND_MESSAGE)
const segment = await payload.find({
collection: 'recordingSegment',
where: {
and: [
{ recording: { equals: parsed.recordingId } },
{ speakerKey: { equals: parsed.speakerKey } },
],
},
depth: 0,
limit: 1,
pagination: false,
select: { speakerKey: true },
req,
user: actor,
overrideAccess: false,
})
if (!segment.docs[0]) throw new Error(RECORDING_SPEAKER_UNKNOWN_MESSAGE)
const labels = [...(recording.speakerLabels ?? [])]
const index = labels.findIndex((entry) => entry.speakerKey === parsed.speakerKey)
const entry = { speakerKey: parsed.speakerKey, label: parsed.label }
if (index >= 0) labels[index] = entry
else labels.push(entry)
await payload.update({
collection: 'recording',
id: parsed.recordingId,
data: { speakerLabels: labels, speakerLabelsDropped: false },
depth: 0,
select: recordingSelect,
req,
user: actor,
overrideAccess: false,
})
})
return { labeled: true }
}
/**
* Hard-deletes one recording: the row and its transcript go in one transaction
* (the collection's `beforeDelete` cascade owns the segments); the private media

View file

@ -1,13 +1,19 @@
import type { CollectionBeforeDeleteHook, CollectionConfig } from 'payload'
import type {
CollectionBeforeDeleteHook,
CollectionBeforeValidateHook,
CollectionConfig,
} from 'payload'
import {
RECORDING_MEDIA_SLUG,
RECORDING_SPEAKER_LABEL_MAX_LENGTH,
RECORDING_STATUSES,
RECORDING_STEPS,
RECORDING_TITLE_MAX_LENGTH,
recordingStatusLabels,
recordingStepLabels,
} from '@/lib/recording'
import { speakerNamesFromLabels } from '@/lib/recordingDiarization'
import {
canDeleteRecording,
canReadRecording,
@ -46,6 +52,21 @@ const deleteRecordingSegments: CollectionBeforeDeleteHook = async ({ id, req })
})
}
/**
* C200 — `speakerNames` is derived from the human labels so the "Pessoa" facet
* filters one denormalized array instead of joining segments. Mirrors
* `RecordingSegment.deriveSearchText`: `originalDoc` covers every update that
* does not touch the labels (status, searchText, step).
*/
const deriveSpeakerNames: CollectionBeforeValidateHook = ({ data, originalDoc }) => {
if (!data) return data
// An explicit array (even empty) is the new truth; only an untouched field
// falls back to the stored document.
const labels = data.speakerLabels === undefined ? originalDoc?.speakerLabels : data.speakerLabels
data.speakerNames = speakerNamesFromLabels(labels)
return data
}
export const Recording: CollectionConfig = {
slug: 'recording',
labels: {
@ -66,6 +87,7 @@ export const Recording: CollectionConfig = {
delete: canDeleteRecording,
},
hooks: {
beforeValidate: [deriveSpeakerNames],
beforeChange: [stampCampaignCreatedBy],
beforeDelete: [deleteRecordingSegments],
},
@ -133,6 +155,52 @@ export const Recording: CollectionConfig = {
description: 'Concatenação normalizada dos segmentos (sem acentos, minúsculas).',
},
},
{
name: 'speakerLabels',
type: 'array',
label: 'Falantes identificados',
admin: {
readOnly: true,
description:
'Rótulos humanos por agrupamento acústico (chave → nome). Identificação sempre humana; o acervo nunca sugere ou infere pessoas.',
},
fields: [
{
name: 'speakerKey',
type: 'text',
label: 'Agrupamento',
required: true,
},
{
name: 'label',
type: 'text',
label: 'Nome do falante',
required: true,
maxLength: RECORDING_SPEAKER_LABEL_MAX_LENGTH,
},
],
},
{
name: 'speakerNames',
type: 'text',
label: 'Pessoas (busca)',
hasMany: true,
admin: {
readOnly: true,
description:
'Derivado dos rótulos (sem duplicatas, sem diferenciar maiúsculas) para a faceta "Pessoa".',
},
},
{
name: 'speakerLabelsDropped',
type: 'checkbox',
label: 'Identificações perdidas no reprocessamento',
admin: {
readOnly: true,
description:
'Ligado quando um reprocessamento reordenou os agrupamentos e algum rótulo não pôde ser re-vinculado.',
},
},
{
name: 'error',
type: 'textarea',

View file

@ -7,7 +7,9 @@ import { canReadRecording, payloadAdminOnly } from '@/utilities/campaignAccess'
* C199 — searchable ASR segment of an uploaded recording: one timestamped block
* of the transcript with the trigram-search index (`recording_segment_search_text_trgm_idx`,
* hand-written migration). Admin-hidden: the product surface is the recording
* detail with its clickable transcript.
* detail with its clickable transcript. C200 adds `speakerKey`: the anonymous
* acoustic cluster ("Falante N") the segment was assigned to; null when the
* recording has no diarization (the detail renders the plain transcript then).
*/
const deriveSearchText: CollectionBeforeValidateHook = ({ data, originalDoc }) => {
@ -52,6 +54,15 @@ export const RecordingSegment: CollectionConfig = {
label: 'Ordem',
required: true,
},
{
name: 'speakerKey',
type: 'text',
label: 'Agrupamento de falante',
admin: {
readOnly: true,
description: 'Chave do agrupamento acústico ("Falante N") do trecho; vazio sem diarização.',
},
},
{
name: 'startSeconds',
type: 'number',

View file

@ -1,14 +1,23 @@
'use client'
import { CircleAlertIcon, DownloadIcon, PlayIcon, UploadIcon } from 'lucide-react'
import { CircleAlertIcon, DownloadIcon, InfoIcon, PlayIcon, UploadIcon } from 'lucide-react'
import { useRef, useState } from 'react'
import { RecordingRetryButton } from '@/components/campaign/recording/RecordingRetryButton'
import { SpeechHighlightParts } from '@/components/campaign/speech/SpeechHighlightParts'
import { RecordingSpeakerTranscript } from '@/components/campaign/recording/RecordingSpeakerTranscript'
import { RecordingTranscriptSegmentButton } from '@/components/campaign/recording/RecordingTranscriptSegmentButton'
import { Button } from '@/components/ui/button'
import { canRetryRecording, type RecordingStatus } from '@/lib/recording'
import {
canRetryRecording,
RECORDING_SPEAKER_INFO_BANNER,
RECORDING_SPEAKER_INFO_BANNER_MOBILE,
type RecordingStatus,
} from '@/lib/recording'
import { cn } from '@/lib/utils'
import type { RecordingDetailSegmentViewModel } from '@/utilities/recordings/recordingViewModels'
import type {
RecordingDetailSegmentViewModel,
RecordingSpeakerGroupViewModel,
} from '@/utilities/recordings/recordingViewModels'
type RecordingDetailPlayerProps = {
recordingId: number
@ -17,6 +26,10 @@ type RecordingDetailPlayerProps = {
downloadHref: string
failureMessage: string | null
segments: readonly RecordingDetailSegmentViewModel[]
/** C200 — groups when the whole transcript is keyed; empty otherwise. */
speakerGroups: readonly RecordingSpeakerGroupViewModel[]
/** C200 — true when a reprocessing dropped some identification. */
speakerLabelsDropped: boolean
/** `?t=` of the search hit; null when the detail was opened without a term. */
initialSeconds: number | null
}
@ -35,6 +48,8 @@ export const RecordingDetailPlayer = ({
downloadHref,
failureMessage,
segments,
speakerGroups,
speakerLabelsDropped,
initialSeconds,
}: RecordingDetailPlayerProps) => {
const videoRef = useRef<HTMLVideoElement>(null)
@ -48,6 +63,7 @@ export const RecordingDetailPlayer = ({
const [retrying, setRetrying] = useState(false)
const hasFile = status !== 'uploading'
const showTranscript = status === 'ready' && segments.length > 0
const grouped = speakerGroups.length > 0
const seekTo = (seconds: number) => {
const video = videoRef.current
@ -157,6 +173,16 @@ export const RecordingDetailPlayer = ({
</p>
) : null}
{grouped ? (
<div className="mt-4 flex gap-2 rounded-lg border border-border bg-muted/50 p-3">
<InfoIcon className="mt-0.5 size-4 shrink-0 text-muted-foreground" aria-hidden="true" />
<p className="text-xs leading-5 text-muted-foreground">
<span className="md:hidden">{RECORDING_SPEAKER_INFO_BANNER_MOBILE}</span>
<span className="max-md:hidden">{RECORDING_SPEAKER_INFO_BANNER}</span>
</p>
</div>
) : null}
{status === 'ready' && segments.length === 0 ? (
<p className="mt-4 text-xs text-muted-foreground">
A transcrição desta gravação está vazia — o arquivo continua disponível para download.
@ -178,48 +204,44 @@ export const RecordingDetailPlayer = ({
return (
<div
data-slot="recording-player"
className="grid gap-6 lg:grid-cols-[minmax(0,1.35fr)_minmax(340px,.65fr)]"
className={cn(
'grid gap-6',
grouped
? 'lg:grid-cols-[minmax(0,1.05fr)_minmax(430px,.95fr)]'
: 'lg:grid-cols-[minmax(0,1.35fr)_minmax(340px,.65fr)]',
)}
>
{mediaColumn}
<section aria-label="Transcrição" className="min-w-0">
<div className="flex items-center justify-between">
<p className="text-xs font-medium tracking-wide text-muted-foreground uppercase">
Transcrição
</p>
<span className="text-[11px] text-muted-foreground">Clique para posicionar</span>
</div>
<ol className="mt-2 max-h-[28rem] space-y-0.5 overflow-y-auto">
{segments.map((segment) => {
const active = activeStart === segment.startSeconds
return (
{grouped ? (
<RecordingSpeakerTranscript
recordingId={recordingId}
groups={speakerGroups}
labelsDropped={speakerLabelsDropped}
activeStart={activeStart}
onSeek={seekTo}
/>
) : (
<section aria-label="Transcrição" className="min-w-0">
<div className="flex items-center justify-between">
<p className="text-xs font-medium tracking-wide text-muted-foreground uppercase">
Transcrição
</p>
<span className="text-[11px] text-muted-foreground">Clique para posicionar</span>
</div>
<ol className="mt-2 max-h-[28rem] space-y-0.5 overflow-y-auto">
{segments.map((segment) => (
<li key={`${segment.startSeconds}-${segment.startLabel}`}>
<button
type="button"
data-start-seconds={segment.startSeconds}
onClick={() => seekTo(segment.startSeconds)}
className={cn(
'grid w-full grid-cols-[3.25rem_1fr] gap-3 rounded-lg px-2 py-1.5 text-left transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring',
active && 'bg-muted',
)}
>
<span
className={cn(
'pt-0.5 text-xs tabular-nums',
active ? 'text-primary' : 'text-muted-foreground',
)}
>
{segment.startLabel}
</span>
<span className="text-sm leading-relaxed text-foreground/90">
<SpeechHighlightParts parts={segment.parts} />
</span>
</button>
<RecordingTranscriptSegmentButton
segment={segment}
active={activeStart === segment.startSeconds}
onSeek={seekTo}
/>
</li>
)
})}
</ol>
</section>
))}
</ol>
</section>
)}
</div>
)
}

View file

@ -0,0 +1,87 @@
'use client'
import { FunnelIcon, XIcon } from 'lucide-react'
import { useOptimistic } from 'react'
import {
CampaignHeaderFilterPopover,
type CampaignHeaderFilterRow,
} from '@/components/campaign/shared/CampaignHeaderFilterPopover'
import { useCampaignListFilterNavigation } from '@/components/campaign/shared/useCampaignListFilterNavigation'
import { foldSpeakerName } from '@/lib/recordingDiarization'
import {
buildRecordingListHref,
removeRecordingPerson,
toggleRecordingPerson,
type RecordingListState,
} from '@/utilities/recordings/recordingListUrl'
const PERSON_FILTER_ID = 'recording-filter-person'
/**
* C200 — the "Pessoa" facet of the uploaded recordings (design scene 3): one
* removable chip per selected label plus the multi-select popover backed by the
* labels actually present. Options come from the server (`speakerNames`), never
* from a free-form directory — the label is text, not a `Contact`.
*/
export const RecordingPersonFilter = ({
state,
options,
}: {
state: RecordingListState
options: readonly string[]
}) => {
const { navigate } = useCampaignListFilterNavigation({
state,
toHref: (next) => buildRecordingListHref(next, 1),
})
const [viewState, setOptimisticState] = useOptimistic(state)
const selected = viewState.people ?? []
const optionRows: CampaignHeaderFilterRow[] = options.map((option) => {
const next = toggleRecordingPerson(viewState, option)
return {
value: option,
label: option,
href: buildRecordingListHref(next, 1),
selected: selected.some((person) => foldSpeakerName(person) === foldSpeakerName(option)),
checkbox: true,
onChoose: () => setOptimisticState(next),
}
})
const clear = {
href: buildRecordingListHref({ ...viewState, page: 1, people: undefined }, 1),
onChoose: () => setOptimisticState({ ...viewState, page: 1, people: undefined }),
}
return (
<div role="group" aria-label="Filtrar por pessoa" className="flex flex-wrap items-center gap-2">
<CampaignHeaderFilterPopover
id={PERSON_FILTER_ID}
label="Pessoa"
triggerVariant="chip"
triggerLabel={selected.length > 1 ? `Pessoa: ${selected.length}` : 'Pessoa'}
active={selected.length > 0}
closeOnChoose={false}
emptyLabel="Nenhuma pessoa identificada nas gravações."
optionRows={optionRows}
clear={selected.length ? clear : undefined}
/>
{selected.map((person) => (
<button
key={person}
type="button"
className="inline-flex h-8 min-h-8 items-center gap-1.5 rounded-lg border border-input bg-white px-3 text-xs font-semibold focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
onClick={() => navigate(removeRecordingPerson(viewState, person))}
aria-label={`Remover filtro de ${person}`}
>
<FunnelIcon className="size-3.5 shrink-0" aria-hidden="true" />
<span className="max-w-56 truncate">Pessoa: {person}</span>
<XIcon className="size-3.5 shrink-0" aria-hidden="true" />
</button>
))}
</div>
)
}

View file

@ -1,4 +1,4 @@
import { ChevronRightIcon, VideoIcon } from 'lucide-react'
import { ChevronRightIcon, UserIcon, VideoIcon } from 'lucide-react'
import Link from 'next/link'
import { RecordingStatusBadge } from '@/components/campaign/recording/RecordingStatusBadge'
@ -65,6 +65,21 @@ export const RecordingResultCard = ({ recording }: { recording: RecordingListIte
{STATE_COPY[recording.status]}
</p>
)}
{recording.matchedPersons.length > 0 ? (
<div className="mt-2 flex flex-wrap items-center gap-2 text-xs">
{recording.matchedPersons.map((person) => (
<span
key={person}
className="inline-flex min-h-7 items-center gap-1.5 rounded-md bg-muted px-2 py-1 text-xs font-semibold"
>
<UserIcon className="size-3.5 shrink-0" aria-hidden="true" />
{person}
</span>
))}
<span className="text-muted-foreground">aparece nesta gravação</span>
</div>
) : null}
</div>
<Button

View file

@ -0,0 +1,174 @@
'use client'
import { InfoIcon, PencilIcon } from 'lucide-react'
import { useRouter } from 'next/navigation'
import { useState } from 'react'
import type { RecordingSpeakerLabelResponse } from '@/app/(campaign)/campanha/(app)/comunicacao/acervo/gravacoes/types'
import { Alert, AlertDescription } from '@/components/ui/Alert'
import { Button } from '@/components/ui/button'
import {
Dialog,
DialogContent,
DialogDescription,
DialogHeader,
DialogTitle,
DialogTrigger,
} from '@/components/ui/dialog'
import { Input } from '@/components/ui/input'
import { Spinner } from '@/components/ui/Spinner'
import { postCampaignJson } from '@/lib/campaignJsonRequest'
import { campaignRecordingSpeakersHref } from '@/lib/campaignPaths'
import {
RECORDING_SPEAKER_DIALOG_HELP,
RECORDING_SPEAKER_DIALOG_INTRO,
RECORDING_SPEAKER_DIALOG_TITLE,
RECORDING_SPEAKER_EDIT_ACTION,
RECORDING_SPEAKER_IDENTIFY_ACTION,
RECORDING_SPEAKER_LABEL_FIELD_LABEL,
RECORDING_SPEAKER_LABEL_LONG_MESSAGE,
RECORDING_SPEAKER_LABEL_MAX_LENGTH,
RECORDING_SPEAKER_LABEL_PLACEHOLDER,
RECORDING_SPEAKER_LABEL_REQUIRED_MESSAGE,
RECORDING_SPEAKER_SAVE_LABEL,
} from '@/lib/recording'
import type { RecordingSpeakerGroupViewModel } from '@/utilities/recordings/recordingViewModels'
type RecordingSpeakerDialogProps = {
recordingId: number
group: RecordingSpeakerGroupViewModel
}
const SAVE_ERROR_MESSAGE = 'Não foi possível salvar a identificação. Tente novamente.'
/**
* C200 — "Identificar falante" (design scene 2): the team names the anonymous
* cluster after listening. The dialog never suggests a person; the label is
* text applied to every segment of the group and the server rejects a key that
* no longer exists in the recording.
*/
export const RecordingSpeakerDialog = ({ recordingId, group }: RecordingSpeakerDialogProps) => {
const router = useRouter()
const [open, setOpen] = useState(false)
const [label, setLabel] = useState(group.label ?? '')
const [error, setError] = useState<string | null>(null)
const [submitting, setSubmitting] = useState(false)
const handleOpenChange = (next: boolean) => {
setOpen(next)
if (next) {
setLabel(group.label ?? '')
setError(null)
}
}
const submit = async () => {
const trimmed = label.trim()
if (!trimmed) {
setError(RECORDING_SPEAKER_LABEL_REQUIRED_MESSAGE)
return
}
if (trimmed.length > RECORDING_SPEAKER_LABEL_MAX_LENGTH) {
setError(RECORDING_SPEAKER_LABEL_LONG_MESSAGE)
return
}
setSubmitting(true)
setError(null)
try {
const { ok, payload } = await postCampaignJson<RecordingSpeakerLabelResponse>(
campaignRecordingSpeakersHref(recordingId),
{ recordingId, speakerKey: group.key, label: trimmed },
)
if (!ok || payload.status !== 'success') {
setError(payload.status === 'error' ? payload.message : SAVE_ERROR_MESSAGE)
return
}
setOpen(false)
router.refresh()
} catch {
setError(SAVE_ERROR_MESSAGE)
} finally {
setSubmitting(false)
}
}
return (
<Dialog open={open} onOpenChange={handleOpenChange}>
<DialogTrigger asChild>
<Button
type="button"
variant="outline"
className="min-h-9 shrink-0 px-2 md:min-h-10 md:px-3"
aria-label={group.isIdentified ? RECORDING_SPEAKER_EDIT_ACTION : undefined}
>
<PencilIcon data-icon="inline-start" aria-hidden="true" />
<span className="hidden md:inline">
{group.isIdentified ? RECORDING_SPEAKER_EDIT_ACTION : RECORDING_SPEAKER_IDENTIFY_ACTION}
</span>
{/* Mobile keeps the short verb; an identified group is icon-only. */}
{group.isIdentified ? null : <span className="md:hidden">Identificar</span>}
</Button>
</DialogTrigger>
<DialogContent className="sm:max-w-md">
<DialogHeader>
<DialogTitle>{RECORDING_SPEAKER_DIALOG_TITLE}</DialogTitle>
<DialogDescription>Dê um nome ao agrupamento “{group.defaultLabel}”.</DialogDescription>
</DialogHeader>
<div className="flex gap-2 rounded-lg border border-border bg-muted/50 p-3">
<InfoIcon className="mt-0.5 size-4 shrink-0 text-muted-foreground" aria-hidden="true" />
<p className="text-xs leading-5 text-muted-foreground">
{RECORDING_SPEAKER_DIALOG_INTRO}
</p>
</div>
<form
className="flex flex-col gap-1.5"
onSubmit={(event) => {
event.preventDefault()
void submit()
}}
>
<label htmlFor={`speaker-label-${group.key}`} className="text-sm font-medium">
{RECORDING_SPEAKER_LABEL_FIELD_LABEL}
</label>
<Input
id={`speaker-label-${group.key}`}
value={label}
onChange={(event) => setLabel(event.target.value)}
maxLength={RECORDING_SPEAKER_LABEL_MAX_LENGTH}
placeholder={RECORDING_SPEAKER_LABEL_PLACEHOLDER}
className="min-h-11 text-sm"
autoFocus
disabled={submitting}
/>
<p className="text-xs text-muted-foreground">{RECORDING_SPEAKER_DIALOG_HELP}</p>
{error ? (
<Alert variant="destructive" className="mt-2 py-2">
<AlertDescription className="text-xs">{error}</AlertDescription>
</Alert>
) : null}
<div className="mt-4 flex flex-col-reverse gap-2 sm:flex-row sm:justify-end">
<Button
type="button"
variant="outline"
className="min-h-11"
onClick={() => setOpen(false)}
disabled={submitting}
>
Cancelar
</Button>
<Button type="submit" className="min-h-11" disabled={submitting}>
{submitting ? <Spinner data-icon="inline-start" aria-hidden="true" /> : null}
{RECORDING_SPEAKER_SAVE_LABEL}
</Button>
</div>
</form>
</DialogContent>
</Dialog>
)
}

View file

@ -0,0 +1,103 @@
'use client'
import { InfoIcon, UserIcon } from 'lucide-react'
import { RecordingSpeakerDialog } from '@/components/campaign/recording/RecordingSpeakerDialog'
import { RecordingTranscriptSegmentButton } from '@/components/campaign/recording/RecordingTranscriptSegmentButton'
import {
RECORDING_SPEAKER_BANNER_BODY,
RECORDING_SPEAKER_BANNER_DROPPED,
RECORDING_SPEAKER_BANNER_TITLE,
RECORDING_SPEAKER_GROUP_IDENTIFIED,
RECORDING_SPEAKER_GROUP_UNIDENTIFIED,
} from '@/lib/recording'
import { cn } from '@/lib/utils'
import type { RecordingSpeakerGroupViewModel } from '@/utilities/recordings/recordingViewModels'
type RecordingSpeakerTranscriptProps = {
recordingId: number
groups: readonly RecordingSpeakerGroupViewModel[]
labelsDropped: boolean
/** `?t=` search hit or the segment under the player's current time. */
activeStart: number | null
onSeek: (seconds: number) => void
}
/**
* C200 — the grouped transcript of the recording detail (design scenes 1, 4 and
* 5): anonymous clusters in first-appearance order, the always-visible
* imprecision warning (the grouping is automatic; the team identifies) and the
* "Identificar falante" affordance per group.
*/
export const RecordingSpeakerTranscript = ({
recordingId,
groups,
labelsDropped,
activeStart,
onSeek,
}: RecordingSpeakerTranscriptProps) => (
<section aria-label="Transcrição por falante" className="min-w-0">
<div className="flex gap-3 rounded-lg border border-amber-200 bg-amber-50 p-4">
<InfoIcon className="mt-0.5 size-4 shrink-0 text-amber-800" aria-hidden="true" />
<div>
<h4 className="text-sm font-semibold text-amber-950">{RECORDING_SPEAKER_BANNER_TITLE}</h4>
<p className="mt-1 text-sm leading-5 text-amber-900">{RECORDING_SPEAKER_BANNER_BODY}</p>
{labelsDropped ? (
<p className="mt-1 text-sm leading-5 text-amber-900">
{RECORDING_SPEAKER_BANNER_DROPPED}
</p>
) : null}
</div>
</div>
<div className="mt-4 flex items-center justify-between">
<h4 className="text-xs font-medium tracking-wide text-muted-foreground uppercase">
Transcrição por falante
</h4>
<span className="text-[11px] text-muted-foreground md:text-xs">
<span className="md:hidden">Toque para posicionar</span>
<span className="max-md:hidden">Clique para posicionar</span>
</span>
</div>
<div className="mt-2 max-h-[28rem] space-y-3 overflow-y-auto">
{groups.map((group) => (
<article key={group.key} className="overflow-hidden rounded-xl border bg-card">
<header className="flex items-center justify-between gap-3 border-b border-border bg-stone-50 px-3 py-2.5">
<div className="flex min-w-0 items-center gap-2">
<span
className={cn(
'grid size-7 shrink-0 place-items-center rounded-full',
group.isIdentified ? 'bg-primary/10 text-primary' : 'bg-muted',
)}
>
<UserIcon className="size-3.5" aria-hidden="true" />
</span>
<div className="min-w-0">
<h5 className="truncate text-sm font-semibold">
{group.label ?? group.defaultLabel}
</h5>
<p className="text-[11px] text-muted-foreground">
{group.isIdentified
? RECORDING_SPEAKER_GROUP_IDENTIFIED
: RECORDING_SPEAKER_GROUP_UNIDENTIFIED}
</p>
</div>
</div>
<RecordingSpeakerDialog recordingId={recordingId} group={group} />
</header>
{group.segments.map((segment) => (
<RecordingTranscriptSegmentButton
key={`${segment.startSeconds}-${segment.startLabel}`}
segment={segment}
active={activeStart === segment.startSeconds}
variant="grouped"
onSeek={onSeek}
/>
))}
</article>
))}
</div>
</section>
)

View file

@ -0,0 +1,51 @@
'use client'
import { SpeechHighlightParts } from '@/components/campaign/speech/SpeechHighlightParts'
import { cn } from '@/lib/utils'
import type { RecordingDetailSegmentViewModel } from '@/utilities/recordings/recordingViewModels'
/**
* C200 — the clickable transcript line shared by the plain C199 list and the
* grouped C200 blocks: seeks the player to the segment and renders the search
* highlight. `variant` keeps the two approved paddings/sizes without a twin.
*/
export const RecordingTranscriptSegmentButton = ({
segment,
active,
variant = 'plain',
onSeek,
}: {
segment: RecordingDetailSegmentViewModel
active: boolean
variant?: 'plain' | 'grouped'
onSeek: (seconds: number) => void
}) => (
<button
type="button"
data-start-seconds={segment.startSeconds}
onClick={() => onSeek(segment.startSeconds)}
className={cn(
'grid w-full grid-cols-[3.25rem_1fr] gap-3 text-left transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring',
variant === 'grouped' ? 'px-3 py-2.5' : 'rounded-lg px-2 py-1.5',
active && 'bg-muted',
)}
>
<span
className={cn(
'text-xs tabular-nums',
variant === 'plain' && 'pt-0.5',
active ? 'text-primary' : 'text-muted-foreground',
)}
>
{segment.startLabel}
</span>
<span
className={cn(
'text-sm text-foreground/90',
variant === 'grouped' ? 'leading-5' : 'leading-relaxed',
)}
>
<SpeechHighlightParts parts={segment.parts} />
</span>
</button>
)

View file

@ -73,4 +73,8 @@ export const campaignRecordingRetryHref = (id: number): string =>
export const campaignRecordingDeleteHref = (id: number): string =>
`${campaignRecordingDetailHref(id)}/apagar`
/** C200 — human identification of one speaker cluster of a recording. */
export const campaignRecordingSpeakersHref = (id: number): string =>
`${campaignRecordingDetailHref(id)}/falantes`
export const CAMPAIGN_UPDATES_HREF = '/campanha/atualizacoes' as const

View file

@ -30,6 +30,11 @@ export const isRecordingStatus = (value: unknown): value is RecordingStatus =>
/** One owner for the persisted field limits (collection, schema and dialog). */
export const RECORDING_TITLE_MAX_LENGTH = 200
/** C200 — the human label the team types for one speaker cluster. */
export const RECORDING_SPEAKER_LABEL_MAX_LENGTH = 120
export const RECORDING_SPEAKER_LABEL_REQUIRED_MESSAGE = 'Informe o nome do falante.'
export const RECORDING_SPEAKER_LABEL_LONG_MESSAGE = `O nome do falante deve ter até ${RECORDING_SPEAKER_LABEL_MAX_LENGTH} caracteres.`
/** Upload ceiling: hours of plenary are the real case, 4 GiB is the guard. */
export const RECORDING_MAX_BYTES = 4 * 1024 * 1024 * 1024
export const RECORDING_MAX_SIZE_LABEL = '4 GB'
@ -90,6 +95,31 @@ export const canServeRecordingMedia = (status: RecordingStatus): boolean => stat
/** The retry button only applies to a failed transcription. */
export const canRetryRecording = (status: RecordingStatus): boolean => status === 'failed'
/** C200 — dialog copy of the human identification step (design scenes 2/5). */
export const RECORDING_SPEAKER_DIALOG_TITLE = 'Identificar falante'
export const RECORDING_SPEAKER_DIALOG_INTRO =
'Ouça um trecho e informe quem é a pessoa. O acervo não sugere nem identifica pessoas automaticamente.'
export const RECORDING_SPEAKER_DIALOG_HELP =
'O nome será aplicado a todas as falas deste agrupamento nesta gravação.'
export const RECORDING_SPEAKER_LABEL_FIELD_LABEL = 'Nome do falante'
export const RECORDING_SPEAKER_LABEL_PLACEHOLDER = 'Digite o nome do falante'
export const RECORDING_SPEAKER_SAVE_LABEL = 'Salvar identificação'
/** C200 — the transcript banner of the grouped detail (design scenes 1/4/5). */
export const RECORDING_SPEAKER_INFO_BANNER =
'Os falantes são agrupados automaticamente pelo áudio. A identificação de cada pessoa é feita pela equipe.'
export const RECORDING_SPEAKER_INFO_BANNER_MOBILE =
'Agrupamento automático e aproximado. A identificação é feita pela equipe.'
export const RECORDING_SPEAKER_BANNER_TITLE = 'A separação por falante pode estar imprecisa'
export const RECORDING_SPEAKER_BANNER_BODY =
'O agrupamento é automático e aproximado. Ruído e falas sobrepostas podem colocar trechos no agrupamento errado. A equipe continua responsável por identificar cada falante.'
export const RECORDING_SPEAKER_BANNER_DROPPED =
'Esta gravação foi reprocessada e algumas identificações não puderam ser mantidas. Confira os falantes.'
export const RECORDING_SPEAKER_GROUP_UNIDENTIFIED = 'Agrupamento sem identificação'
export const RECORDING_SPEAKER_GROUP_IDENTIFIED = 'Identificação feita pela equipe'
export const RECORDING_SPEAKER_IDENTIFY_ACTION = 'Identificar falante'
export const RECORDING_SPEAKER_EDIT_ACTION = 'Editar identificação'
/** Errors of the upload form (client validation). Wire messages live in schemas. */
export const RECORDING_FILE_REQUIRED_MESSAGE = 'Escolha um arquivo de vídeo para enviar.'
export const RECORDING_TITLE_REQUIRED_MESSAGE = 'Informe um título para a gravação.'

View file

@ -0,0 +1,241 @@
/**
* C200 — pure speaker-grouping rules of an uploaded recording. The diarization
* provider returns anonymous turns (acoustic clusters); this module assigns
* each transcript segment to the turn it overlaps and numbers the clusters
* globally by first appearance. Identification is human: nothing here matches a
* voice to a person — the labels are text the team types.
*
* No I/O and no `server-only`: the job, the view models and the unit tests
* share this module. `reconcileSpeakerLabels` is the reprocessing contract: a
* new run cannot re-associate a voice (no biometrics), so a label survives only
* when an old cluster maps 1:1 to a new one by time overlap; the rest is
* reported as dropped so the detail can warn instead of silently forgetting.
*/
/** One anonymous speaker turn as the provider returns it, in absolute seconds. */
export type DiarizedTurn = {
speaker: string
startSeconds: number
endSeconds: number
}
/** A segment with the cluster key it was assigned to; null if no turn covers it. */
export type SpeakerKeyed = {
speakerKey: string | null
}
/** A label the team typed for one cluster key of one recording. */
export type RecordingSpeakerLabel = {
speakerKey: string
label: string
}
/** A span used by the alignment/reconciliation math (segment or cluster). */
export type TimedSpan = {
startSeconds: number
endSeconds: number
}
/**
* Overlap in seconds between two spans; zero when they do not touch. A
* degenerate span (end before start) is clamped to its start, never negative.
*/
const overlapSeconds = (a: TimedSpan, b: TimedSpan): number =>
Math.max(0, Math.min(a.endSeconds, b.endSeconds) - Math.max(a.startSeconds, b.startSeconds))
/** Gap in seconds between two spans; zero when they overlap or touch. */
const gapSeconds = (a: TimedSpan, b: TimedSpan): number =>
Math.max(0, Math.max(a.startSeconds, b.startSeconds) - Math.min(a.endSeconds, b.endSeconds))
/** Closest boundary distance used to break overlap ties by midpoint proximity. */
const midpointDistance = (a: TimedSpan, b: TimedSpan): number =>
Math.abs((a.startSeconds + a.endSeconds) / 2 - (b.startSeconds + b.endSeconds) / 2)
/**
* Picks the turn of one segment: the largest overlap wins, ties go to the turn
* whose midpoint sits closest to the segment's, and a segment that overlaps no
* turn at all falls back to the nearest turn by gap (the audio is contiguous,
* so an isolated segment belongs to the closest voice). Invalid turns are
* ignored by the caller.
*/
const pickTurn = (segment: TimedSpan, turns: readonly DiarizedTurn[]): number => {
let best = -1
let bestOverlap = 0
let bestDistance = Number.POSITIVE_INFINITY
for (const [turnIndex, turn] of turns.entries()) {
const overlap = overlapSeconds(segment, turn)
const distance = midpointDistance(segment, turn)
const wins =
overlap > bestOverlap || (overlap > 0 && overlap === bestOverlap && distance < bestDistance)
if (wins) {
best = turnIndex
bestOverlap = overlap
bestDistance = distance
}
}
if (bestOverlap > 0) return best
let nearest = 0
let nearestGap = Number.POSITIVE_INFINITY
for (const [turnIndex, turn] of turns.entries()) {
const gap = gapSeconds(segment, turn)
if (gap < nearestGap) {
nearest = turnIndex
nearestGap = gap
}
}
return nearest
}
/**
* Assigns each segment a `speaker-N` key following the acoustic turns. The
* numbering is global to the recording and follows the first appearance of the
* cluster in segment order — two runs over the same audio produce the same
* keys, and a segment with no usable turn keeps a null key (the detail renders
* the plain transcript then). Pure: tests do not touch the provider.
*/
export const assignSpeakerKeys = <Segment extends TimedSpan>(
segments: readonly Segment[],
turns: readonly DiarizedTurn[],
): (Segment & SpeakerKeyed)[] => {
const validTurns = turns.filter(
(turn) =>
typeof turn.speaker === 'string' &&
turn.speaker.length > 0 &&
Number.isFinite(turn.startSeconds) &&
Number.isFinite(turn.endSeconds),
)
if (validTurns.length === 0) {
return segments.map((segment) => ({ ...segment, speakerKey: null }))
}
const clusterKeys = new Map<string, string>()
const keyOfTurn = (turn: DiarizedTurn): string => {
const existing = clusterKeys.get(turn.speaker)
if (existing) return existing
const key = `speaker-${clusterKeys.size + 1}`
clusterKeys.set(turn.speaker, key)
return key
}
return segments.map((segment) => {
const turn = validTurns[pickTurn(segment, validTurns)]
return { ...segment, speakerKey: turn ? keyOfTurn(turn) : null }
})
}
/** First-appearance order of the keys in a keyed segment list. */
export const speakerKeysInOrder = (segments: readonly SpeakerKeyed[]): string[] => {
const keys: string[] = []
const seen = new Set<string>()
for (const segment of segments) {
if (!segment.speakerKey || seen.has(segment.speakerKey)) continue
seen.add(segment.speakerKey)
keys.push(segment.speakerKey)
}
return keys
}
/**
* Reprocessing contract (D6): a previous label is kept only on a mutual 1:1
* time-overlap match — one old cluster maps to exactly one new cluster and that
* new cluster receives exactly one old label. A split (one old cluster became
* two) or a merge (two old clusters became one) is uncertainty, so the affected
* labels are dropped and reported instead of naming the wrong voice. Output
* labels follow the new segment order.
*/
export const reconcileSpeakerLabels = ({
previousLabels,
previousSegments,
nextSegments,
}: {
previousLabels: readonly RecordingSpeakerLabel[]
previousSegments: readonly (TimedSpan & SpeakerKeyed)[]
nextSegments: readonly (TimedSpan & SpeakerKeyed)[]
}): { labels: RecordingSpeakerLabel[]; dropped: boolean } => {
if (previousLabels.length === 0) return { labels: [], dropped: false }
const nextKeys = speakerKeysInOrder(nextSegments)
if (nextKeys.length === 0) return { labels: [], dropped: true }
const previousKeys = speakerKeysInOrder(previousSegments)
const overlapByPair = new Map<string, number>()
for (const previous of previousSegments) {
if (!previous.speakerKey) continue
for (const next of nextSegments) {
if (!next.speakerKey) continue
const pair = `${previous.speakerKey}\u0000${next.speakerKey}`
overlapByPair.set(pair, (overlapByPair.get(pair) ?? 0) + overlapSeconds(previous, next))
}
}
const labelByPreviousKey = new Map(previousLabels.map((entry) => [entry.speakerKey, entry.label]))
const claimantsOfNextKey = (nextKey: string): string[] =>
previousKeys.filter(
(previousKey) => (overlapByPair.get(`${previousKey}\u0000${nextKey}`) ?? 0) > 0,
)
const kept = new Map<string, string>()
for (const nextKey of nextKeys) {
const claimants = claimantsOfNextKey(nextKey)
const [previousKey] = claimants
if (claimants.length !== 1 || !previousKey) continue
const label = labelByPreviousKey.get(previousKey)
if (!label) continue
// Mutual uniqueness: the same old cluster must not also map elsewhere.
const matchesOfPrevious = nextKeys.filter(
(candidate) => (overlapByPair.get(`${previousKey}\u0000${candidate}`) ?? 0) > 0,
)
if (matchesOfPrevious.length !== 1) continue
kept.set(nextKey, label)
}
const labels = [...kept.entries()].map(([speakerKey, label]) => ({ speakerKey, label }))
return { labels, dropped: previousLabels.length > labels.length }
}
/**
* Distinct labels of a recording, case-insensitive and trimmed — the
* denormalized `speakerNames` used by the "Pessoa" facet. Order of first
* appearance; empty labels never count.
*/
export const speakerNamesFromLabels = (
labels: readonly RecordingSpeakerLabel[] | null | undefined,
): string[] => {
const names: string[] = []
const seen = new Set<string>()
for (const entry of labels ?? []) {
const name = typeof entry?.label === 'string' ? entry.label.trim() : ''
if (!name) continue
const folded = foldSpeakerName(name)
if (seen.has(folded)) continue
seen.add(folded)
names.push(name)
}
return names
}
/** The default public name of one cluster: always numbered, never a person. */
export const recordingSpeakerDefaultLabel = (index: number): string => `Falante ${index}`
/**
* Case-fold used by every speaker-name comparison: the facet query is
* case-insensitive (ILIKE), so parsing, dedupe and chip matching share ONE
* fold to never disagree.
*/
export const foldSpeakerName = (value: string): string => value.toLocaleLowerCase('pt-BR')
/**
* Case-insensitive containment of one selected "Pessoa" term in one label —
* the same match the facet query runs (`contains` → ILIKE), so the card chip
* and the result set never disagree.
*/
export const speakerNameMatches = (name: string, filter: string): boolean =>
foldSpeakerName(name).includes(foldSpeakerName(filter))
/**
* Provider ceiling for one diarization call (AssemblyAI accepts up to 10 h per
* file). Longer recordings skip the step and keep the honest single-block
* transcript instead of guessing speakers.
*/
export const RECORDING_DIARIZATION_MAX_SECONDS = 10 * 60 * 60

View file

@ -34,17 +34,11 @@ export type RecordingTranscriptSegment = {
}
/**
* Extracts a mono 16 kHz MP3 and splits it into fixed chunks in one ffmpeg
* pass. Args are returned as an array — never a shell string; `outputPattern`
* carries the `%03d` sequence ffmpeg replaces per chunk.
* The shared extraction prefix of both ffmpeg passes: mono 16 kHz MP3, at the
* bitrate the ASR provider and the diarization provider both accept. Args are
* returned as an array — never a shell string.
*/
export const buildRecordingAudioFfmpegArgs = ({
inputPath,
outputPattern,
}: {
inputPath: string
outputPattern: string
}): string[] => [
const recordingAudioArgs = (inputPath: string): string[] => [
'-nostdin',
'-hide_banner',
'-y',
@ -59,6 +53,20 @@ export const buildRecordingAudioFfmpegArgs = ({
'libmp3lame',
'-b:a',
RECORDING_AUDIO_BITRATE,
]
/**
* Extracts a mono 16 kHz MP3 and splits it into fixed chunks in one ffmpeg
* pass. `outputPattern` carries the `%03d` sequence ffmpeg replaces per chunk.
*/
export const buildRecordingAudioFfmpegArgs = ({
inputPath,
outputPattern,
}: {
inputPath: string
outputPattern: string
}): string[] => [
...recordingAudioArgs(inputPath),
'-f',
'segment',
'-segment_time',
@ -68,6 +76,20 @@ export const buildRecordingAudioFfmpegArgs = ({
outputPattern,
]
/**
* Extracts the whole recording as one mono 16 kHz MP3 (C200): the diarization
* provider needs the entire audio in a single call so the speaker numbering is
* global — a per-chunk run would renumber speakers that never relate across
* chunks. Same audio parameters as the chunked pass.
*/
export const buildRecordingFullAudioFfmpegArgs = ({
inputPath,
outputPath,
}: {
inputPath: string
outputPath: string
}): string[] => [...recordingAudioArgs(inputPath), outputPath]
/**
* Merges every chunk into one ordered transcript: each segment is shifted by
* its chunk offset, empty text is dropped and a degenerate span (end before

View file

@ -1,6 +1,12 @@
import { z } from 'zod'
import { RECORDING_TITLE_MAX_LENGTH, RECORDING_TITLE_REQUIRED_MESSAGE } from '@/lib/recording'
import {
RECORDING_SPEAKER_LABEL_LONG_MESSAGE,
RECORDING_SPEAKER_LABEL_MAX_LENGTH,
RECORDING_SPEAKER_LABEL_REQUIRED_MESSAGE,
RECORDING_TITLE_MAX_LENGTH,
RECORDING_TITLE_REQUIRED_MESSAGE,
} from '@/lib/recording'
import { trimmedOptionalText } from '@/lib/schemas/primitives'
/** The acervo gate refused the actor — one literal shared with the vertical. */
@ -49,6 +55,23 @@ export const recordingRetryRequestSchema = z.object({
recordingId: z.number().int().positive(),
})
/** C200 — the team names one acoustic cluster of a recording. */
export const RECORDING_SPEAKER_UNKNOWN_MESSAGE =
'Este agrupamento de falantes não existe mais nesta gravação.'
export const recordingSpeakerLabelRequestSchema = z.object({
recordingId: z.number().int().positive(),
speakerKey: z
.string()
.trim()
.regex(/^speaker-\d{1,3}$/, RECORDING_SPEAKER_UNKNOWN_MESSAGE),
label: z
.string()
.trim()
.min(1, RECORDING_SPEAKER_LABEL_REQUIRED_MESSAGE)
.max(RECORDING_SPEAKER_LABEL_MAX_LENGTH, RECORDING_SPEAKER_LABEL_LONG_MESSAGE),
})
/** Delete one recording (row + segments + private media). */
export const recordingDeleteRequestSchema = z.object({
recordingId: z.number().int().positive(),

File diff suppressed because it is too large Load diff

View file

@ -0,0 +1,51 @@
import { MigrateUpArgs, MigrateDownArgs, sql } from '@payloadcms/db-postgres'
/**
* C200 — speaker grouping of the uploaded recordings: the anonymous cluster key
* on each segment, the human labels on the recording (one row per identified
* cluster) and the flag a reprocessing sets when a label could not be
* re-associated. `recording_texts` is Payload's join table for the derived
* `speakerNames` (`hasMany text`).
*
* The generated diff also carried `reel.source_hash` because the C199 snapshot
* (20260919_003440) was taken before the sibling `20260919_002605` landed — the
* column already exists in every environment, so that SQL was removed here by
* hand. The snapshot JSON of THIS migration is current, so the phantom diff
* does not come back.
*/
export async function up({ db, payload, req }: MigrateUpArgs): Promise<void> {
await db.execute(sql`
CREATE TABLE "recording_speaker_labels" (
"_order" integer NOT NULL,
"_parent_id" integer NOT NULL,
"id" varchar PRIMARY KEY NOT NULL,
"speaker_key" varchar NOT NULL,
"label" varchar NOT NULL
);
CREATE TABLE "recording_texts" (
"id" serial PRIMARY KEY NOT NULL,
"order" integer NOT NULL,
"parent_id" integer NOT NULL,
"path" varchar NOT NULL,
"text" varchar
);
ALTER TABLE "recording" ADD COLUMN "speaker_labels_dropped" boolean;
ALTER TABLE "recording_segment" ADD COLUMN "speaker_key" varchar;
ALTER TABLE "recording_speaker_labels" ADD CONSTRAINT "recording_speaker_labels_parent_id_fk" FOREIGN KEY ("_parent_id") REFERENCES "public"."recording"("id") ON DELETE cascade ON UPDATE no action;
ALTER TABLE "recording_texts" ADD CONSTRAINT "recording_texts_parent_fk" FOREIGN KEY ("parent_id") REFERENCES "public"."recording"("id") ON DELETE cascade ON UPDATE no action;
CREATE INDEX "recording_speaker_labels_order_idx" ON "recording_speaker_labels" USING btree ("_order");
CREATE INDEX "recording_speaker_labels_parent_id_idx" ON "recording_speaker_labels" USING btree ("_parent_id");
CREATE INDEX "recording_texts_order_parent" ON "recording_texts" USING btree ("order","parent_id");`)
}
export async function down({ db, payload, req }: MigrateDownArgs): Promise<void> {
await db.execute(sql`
ALTER TABLE "recording_speaker_labels" DISABLE ROW LEVEL SECURITY;
ALTER TABLE "recording_texts" DISABLE ROW LEVEL SECURITY;
DROP TABLE "recording_speaker_labels" CASCADE;
DROP TABLE "recording_texts" CASCADE;
ALTER TABLE "recording" DROP COLUMN "speaker_labels_dropped";
ALTER TABLE "recording_segment" DROP COLUMN "speaker_key";`)
}

View file

@ -75,6 +75,7 @@ import * as migration_20260915_142058_activity_google_import_link from './202609
import * as migration_20260918_201009_add_reel from './20260918_201009_add_reel';
import * as migration_20260919_002605_add_reel_source_hash from './20260919_002605_add_reel_source_hash';
import * as migration_20260919_003440_add_recording from './20260919_003440_add_recording';
import * as migration_20260919_032729_add_recording_speakers from './20260919_032729_add_recording_speakers';
export const migrations = [
{
@ -455,11 +456,16 @@ export const migrations = [
{
up: migration_20260919_002605_add_reel_source_hash.up,
down: migration_20260919_002605_add_reel_source_hash.down,
name: '20260919_002605_add_reel_source_hash'
name: '20260919_002605_add_reel_source_hash',
},
{
up: migration_20260919_003440_add_recording.up,
down: migration_20260919_003440_add_recording.down,
name: '20260919_003440_add_recording'
name: '20260919_003440_add_recording',
},
{
up: migration_20260919_032729_add_recording_speakers.up,
down: migration_20260919_032729_add_recording_speakers.down,
name: '20260919_032729_add_recording_speakers'
},
];

View file

@ -1175,6 +1175,24 @@ export interface Recording {
* Concatenação normalizada dos segmentos (sem acentos, minúsculas).
*/
searchText?: string | null;
/**
* Rótulos humanos por agrupamento acústico (chave → nome). Identificação sempre humana; o acervo nunca sugere ou infere pessoas.
*/
speakerLabels?:
| {
speakerKey: string;
label: string;
id?: string | null;
}[]
| null;
/**
* Derivado dos rótulos (sem duplicatas, sem diferenciar maiúsculas) para a faceta "Pessoa".
*/
speakerNames?: string[] | null;
/**
* Ligado quando um reprocessamento reordenou os agrupamentos e algum rótulo não pôde ser re-vinculado.
*/
speakerLabelsDropped?: boolean | null;
/**
* Motivo interno da falha; nunca vai à pessoa com o detalhe cru.
*/
@ -1215,6 +1233,10 @@ export interface RecordingSegment {
id: number;
recording: number | Recording;
order: number;
/**
* Chave do agrupamento acústico ("Falante N") do trecho; vazio sem diarização.
*/
speakerKey?: string | null;
startSeconds: number;
endSeconds: number;
text: string;
@ -2425,6 +2447,15 @@ export interface RecordingSelect<T extends boolean = true> {
media?: T;
durationSeconds?: T;
searchText?: T;
speakerLabels?:
| T
| {
speakerKey?: T;
label?: T;
id?: T;
};
speakerNames?: T;
speakerLabelsDropped?: T;
error?: T;
createdBy?: T;
updatedAt?: T;
@ -2455,6 +2486,7 @@ export interface RecordingMediaSelect<T extends boolean = true> {
export interface RecordingSegmentSelect<T extends boolean = true> {
recording?: T;
order?: T;
speakerKey?: T;
startSeconds?: T;
endSeconds?: T;
text?: T;

View file

@ -0,0 +1,178 @@
import 'server-only'
import { openAsBlob } from 'node:fs'
import type { DiarizedTurn } from '@/lib/recordingDiarization'
/**
* C200 — the diarization provider: AssemblyAI's async pre-recorded API with
* `speaker_labels` (acoustic clustering only). The recording is sent as ONE
* file so the speaker numbering is global; the voice is never matched to a
* person and the account-level speaker identification feature is deliberately
* never enabled — naming is human (design scenes 1/2).
*
* Never throws: every failure becomes `{ ok: false }` and the transcription job
* degrades to the honest single-block transcript (no speaker grouping) instead
* of failing the recording. The remote transcript/audio is deleted best-effort
* after the read.
*
* Provider notes (verified 2026-09-19): Universal-2 at US$0,15/h + async
* diarization add-on US$0,02/h; up to 10 h / 5 GB per file; `speech_models` is
* the plural request field (the singular is deprecated).
*/
export const ASSEMBLYAI_UPLOAD_URL = 'https://api.assemblyai.com/v2/upload'
export const ASSEMBLYAI_TRANSCRIPT_URL = 'https://api.assemblyai.com/v2/transcript'
const ASSEMBLYAI_MODEL = 'universal-2'
const ASSEMBLYAI_LANGUAGE = 'pt'
const UPLOAD_TIMEOUT_MS = 10 * 60_000
const POLL_REQUEST_TIMEOUT_MS = 30_000
const DIARIZATION_DEADLINE_MS = 60 * 60_000
const POLL_INTERVAL_MS = 5_000
const DIARIZE_ERROR = 'Não foi possível separar os falantes da gravação.'
type DiarizeRequestOptions = {
/** Called after each poll — lets the job keep its heartbeat alive. */
onProgress?: () => Promise<void> | void
/** Poll cadence; only tests change it (default 5 s). */
pollIntervalMs?: number
}
type DiarizeResult =
| { ok: true; turns: DiarizedTurn[] }
| { ok: false; error: string; status: number }
/** The seam the transcription job consumes; tests inject a fake. */
export type SpeakerDiarizer = (
file: Blob,
options?: DiarizeRequestOptions,
) => Promise<DiarizeResult>
const sleep = (ms: number): Promise<void> => new Promise((resolve) => setTimeout(resolve, ms))
const authHeaders = (apiKey: string, json = false): Record<string, string> => ({
authorization: apiKey,
...(json ? { 'content-type': 'application/json' } : {}),
})
/** `A`→`speaker-A`… provider cluster ids stay opaque; the pure module numbers them. */
const normalizeUtterances = (json: unknown): DiarizedTurn[] => {
const data = json && typeof json === 'object' ? (json as Record<string, unknown>) : {}
const utterances = Array.isArray(data.utterances) ? data.utterances : []
const turns: DiarizedTurn[] = []
for (const raw of utterances) {
const utterance = raw && typeof raw === 'object' ? (raw as Record<string, unknown>) : {}
const speaker = typeof utterance.speaker === 'string' ? utterance.speaker.trim() : ''
const startMs = Number(utterance.start)
const endMs = Number(utterance.end)
if (!speaker || !Number.isFinite(startMs) || !Number.isFinite(endMs)) continue
turns.push({
speaker,
startSeconds: Math.max(0, startMs / 1000),
endSeconds: Math.max(0, endMs / 1000),
})
}
return turns
}
/** Best-effort delete of the remote transcript (privacy hygiene, never fatal). */
const deleteRemoteTranscript = async (apiKey: string, transcriptId: string): Promise<void> => {
try {
await fetch(`${ASSEMBLYAI_TRANSCRIPT_URL}/${transcriptId}`, {
method: 'DELETE',
headers: authHeaders(apiKey),
signal: AbortSignal.timeout(POLL_REQUEST_TIMEOUT_MS),
})
} catch {
// Best-effort: a failed delete never affects the recording.
}
}
/**
* Uploads the audio, submits the diarization transcript and polls until it is
* ready. `file` is a Blob opened from the extracted MP3 (the job owns the temp
* file); the provider call itself keeps no copy after the best-effort delete.
*/
export const assemblyAiDiarize: SpeakerDiarizer = async (
file,
{ onProgress, pollIntervalMs = POLL_INTERVAL_MS } = {},
) => {
const apiKey = process.env.ASSEMBLYAI_API_KEY
if (!apiKey) {
return { ok: false, error: 'Separação de falantes está indisponível no momento.', status: 503 }
}
let transcriptId: string | null = null
try {
const uploadResponse = await fetch(ASSEMBLYAI_UPLOAD_URL, {
method: 'POST',
headers: authHeaders(apiKey),
body: file,
signal: AbortSignal.timeout(UPLOAD_TIMEOUT_MS),
})
if (!uploadResponse.ok) return { ok: false, error: DIARIZE_ERROR, status: 502 }
const uploaded = (await uploadResponse.json()) as { upload_url?: unknown }
const uploadUrl = typeof uploaded.upload_url === 'string' ? uploaded.upload_url : ''
if (!uploadUrl) return { ok: false, error: DIARIZE_ERROR, status: 502 }
const createResponse = await fetch(ASSEMBLYAI_TRANSCRIPT_URL, {
method: 'POST',
headers: authHeaders(apiKey, true),
body: JSON.stringify({
audio_url: uploadUrl,
language_code: ASSEMBLYAI_LANGUAGE,
speaker_labels: true,
speech_models: [ASSEMBLYAI_MODEL],
}),
signal: AbortSignal.timeout(POLL_REQUEST_TIMEOUT_MS),
})
if (!createResponse.ok) return { ok: false, error: DIARIZE_ERROR, status: 502 }
const created = (await createResponse.json()) as { id?: unknown; status?: unknown }
transcriptId = typeof created.id === 'string' ? created.id : null
if (!transcriptId) return { ok: false, error: DIARIZE_ERROR, status: 502 }
const deadline = Date.now() + DIARIZATION_DEADLINE_MS
for (;;) {
await sleep(pollIntervalMs)
await onProgress?.()
const pollResponse = await fetch(`${ASSEMBLYAI_TRANSCRIPT_URL}/${transcriptId}`, {
headers: authHeaders(apiKey),
signal: AbortSignal.timeout(POLL_REQUEST_TIMEOUT_MS),
})
if (!pollResponse.ok) return { ok: false, error: DIARIZE_ERROR, status: 502 }
const transcript = (await pollResponse.json()) as { status?: unknown }
if (transcript.status === 'completed') {
await deleteRemoteTranscript(apiKey, transcriptId)
return { ok: true, turns: normalizeUtterances(transcript) }
}
if (transcript.status === 'error') {
await deleteRemoteTranscript(apiKey, transcriptId)
return { ok: false, error: DIARIZE_ERROR, status: 502 }
}
if (Date.now() > deadline) {
await deleteRemoteTranscript(apiKey, transcriptId)
return { ok: false, error: DIARIZE_ERROR, status: 504 }
}
}
} catch {
if (transcriptId) await deleteRemoteTranscript(apiKey, transcriptId)
return { ok: false, error: DIARIZE_ERROR, status: 502 }
}
}
/**
* The configured provider, or null when the key is absent. The job treats null
* as "no diarization step": the transcript is still saved, without groups.
*/
export const configuredSpeakerDiarizer = (): SpeakerDiarizer | null =>
process.env.ASSEMBLYAI_API_KEY ? assemblyAiDiarize : null
/** Opens the extracted MP3 as a Blob without loading hours of audio in memory. */
export const diarizeInputBlob = (path: string): Promise<Blob> =>
openAsBlob(path, { type: 'audio/mpeg' })

View file

@ -10,13 +10,24 @@ import {
RECORDING_MEDIA_SLUG,
type RecordingStep,
} from '@/lib/recording'
import {
assignSpeakerKeys,
reconcileSpeakerLabels,
RECORDING_DIARIZATION_MAX_SECONDS,
} from '@/lib/recordingDiarization'
import {
buildRecordingAudioFfmpegArgs,
buildRecordingFullAudioFfmpegArgs,
mergeChunkTranscriptions,
RECORDING_AUDIO_CHUNK_SECONDS,
recordingSearchText,
} from '@/lib/recordingTranscription'
import type { Recording } from '@/payload-types'
import {
configuredSpeakerDiarizer,
diarizeInputBlob,
type SpeakerDiarizer,
} from '@/utilities/ai/assemblyAiDiarize'
import {
deepInfraTranscribeSegments,
type TranscribeSegmentsResult,
@ -107,12 +118,16 @@ const chunkFilesIn = async (tempDir: string): Promise<string[]> => {
/**
* The whole pipeline of one recording. Never throws: an orphan job (server
* restart) is covered by the lazy reaper, and any error marks the row `failed`
* so the retry reuses it.
* so the retry reuses it. C200 adds the optional diarization pass: with a
* configured provider the whole extracted audio is sent once, the turns are
* aligned to the whisper segments and the human labels are reconciled by time
* overlap; a provider failure degrades to the single-block transcript.
*/
export const runRecordingJob = async (
payload: Payload,
recordingId: number,
transcribe: ChunkTranscriber = deepInfraTranscribeSegments,
diarize: SpeakerDiarizer | null = configuredSpeakerDiarizer(),
): Promise<void> => {
let tempDir: string | null = null
let currentStep: RecordingStep = 'extracting'
@ -155,6 +170,22 @@ export const runRecordingJob = async (
return
}
// C200: the provider accepts up to 10 h per file; a longer recording skips
// the step honestly (a rough upper bound is enough to refuse, and the
// chunk count overestimates the real duration by at most one chunk).
const diarizationEligible =
diarize !== null &&
chunkFiles.length * RECORDING_AUDIO_CHUNK_SECONDS <= RECORDING_DIARIZATION_MAX_SECONDS
const fullAudioPath = join(tempDir, 'full.mp3')
if (diarizationEligible) {
await runFfmpeg(
buildRecordingFullAudioFfmpegArgs({ inputPath, outputPath: fullAudioPath }),
RECORDING_AUDIO_CHUNK_SECONDS,
NO_AUDIO_MESSAGE,
RECORDING_FFMPEG_TIMEOUT_MS,
)
}
await markStep('transcribing')
const chunks: {
offsetSeconds: number
@ -187,8 +218,63 @@ export const runRecordingJob = async (
return
}
// C200 — optional speaker grouping. The provider seam never fails the
// recording: every failure (or a missing key) keeps the plain transcript.
let keyedSegments: ((typeof segments)[number] & { speakerKey: string | null })[] = segments.map(
(segment) => ({ ...segment, speakerKey: null }),
)
if (diarizationEligible && diarize) {
try {
const result = await diarize(await diarizeInputBlob(fullAudioPath), {
onProgress: async () => {
await updateRecordingSystem(payload, recordingId, { step: 'transcribing' })
},
})
if (result.ok) keyedSegments = assignSpeakerKeys(segments, result.turns)
} catch {
// Defensive: the provider contract is "never throws"; if it does, the
// transcript still gets saved without groups.
}
}
await markStep('saving')
await withPayloadTransaction(payload, async ({ req }) => {
// Fresh reads inside the transaction: the labels a curator wrote while
// the transcription ran must be the ones reconciled, not the snapshot
// from the top of the job.
const freshRecording = await payload.find({
collection: 'recording',
where: { id: { equals: recordingId } },
depth: 0,
limit: 1,
pagination: false,
select: { speakerLabels: true },
req,
// Intentional admin bypass: the pipeline owns the row state.
overrideAccess: true,
})
const previous = await payload.find({
collection: 'recordingSegment',
where: { recording: { equals: recordingId } },
depth: 0,
limit: 0,
pagination: false,
sort: 'order',
select: { speakerKey: true, startSeconds: true, endSeconds: true },
req,
// Intentional admin bypass: reconciling the transcript the job owns.
overrideAccess: true,
})
const { labels, dropped } = reconcileSpeakerLabels({
previousLabels: freshRecording.docs[0]?.speakerLabels ?? [],
previousSegments: previous.docs.map((segment) => ({
speakerKey: segment.speakerKey ?? null,
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
})),
nextSegments: keyedSegments,
})
await payload.delete({
collection: 'recordingSegment',
where: { recording: { equals: recordingId } },
@ -196,7 +282,7 @@ export const runRecordingJob = async (
// Intentional admin bypass: replacing segments owned by the recording.
overrideAccess: true,
})
for (const segment of segments) {
for (const segment of keyedSegments) {
await payload.create({
collection: 'recordingSegment',
// `searchText` is filled by the collection's own beforeValidate hook.
@ -206,6 +292,7 @@ export const runRecordingJob = async (
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
text: segment.text,
speakerKey: segment.speakerKey,
}),
req,
// Intentional admin bypass: the transcript belongs to the recording row.
@ -221,6 +308,8 @@ export const runRecordingJob = async (
error: null,
durationSeconds: Math.round(durationSeconds),
searchText: recordingSearchText(segments),
speakerLabels: labels,
speakerLabelsDropped: dropped,
},
req,
)

View file

@ -1,16 +1,29 @@
/**
* C199 — Payload `where` of the uploaded-recordings list: one textual branch
* over the normalized `recording.searchText` (the concatenation of the
* transcript segments). No facets: the Câmara facets do not apply to a
* recording. A recording still processing has no `searchText` and is honestly
* out of the search until the transcript is saved.
* transcript segments). C200 adds the "Pessoa" facet: any selected label that
* appears in the recording's derived `speakerNames`. A recording still
* processing has no `searchText` and is honestly out of the search until the
* transcript is saved.
*/
import type { Where } from 'payload'
import { normalizeForSearch } from '@/lib/speechSearch'
import { collapseListWhereOrBranches } from '@/utilities/campaignListUrl'
import type { RecordingListState } from '@/utilities/recordings/recordingListUrl'
export const buildRecordingListWhere = (state: RecordingListState): Where => {
const branches: Where[] = []
const normalized = normalizeForSearch(state.q ?? '')
return normalized ? { searchText: { like: normalized } } : {}
if (normalized) branches.push({ searchText: { like: normalized } })
const peopleBranch = collapseListWhereOrBranches(
(state.people ?? []).map((person) => ({ speakerNames: { contains: person } })),
)
if (peopleBranch) branches.push(peopleBranch)
if (branches.length === 0) return {}
if (branches.length === 1) return branches[0] ?? {}
return { and: branches }
}

View file

@ -6,7 +6,10 @@
*/
import { CAMPAIGN_COMMUNICATION_ACERVO } from '@/lib/campaignPaths'
import { isContactSearchQueryReady } from '@/lib/contactSearchQuery'
import { RECORDING_SPEAKER_LABEL_MAX_LENGTH } from '@/lib/recording'
import { foldSpeakerName } from '@/lib/recordingDiarization'
import {
allParamValues,
buildListHref,
firstValue,
normalizedText,
@ -22,29 +25,52 @@ export type AcervoSource = 'camara' | 'enviadas'
const ACERVO_SOURCE_PARAM = 'source'
const ACERVO_SOURCE_ENVIADAS = 'enviadas'
const ACERVO_PERSON_PARAM = 'person'
const recordingListParamNames = [ACERVO_SOURCE_PARAM, 'q', 'page'] as const
const recordingListParamNames = [ACERVO_SOURCE_PARAM, 'q', 'page', ACERVO_PERSON_PARAM] as const
const recordingListParamNameSet = new Set<string>(recordingListParamNames)
export type RecordingListState = {
source: 'enviadas'
page: number
q?: string
/** Selected "Pessoa" labels; repeated `person` params, literal text. */
people?: string[]
}
/** `enviadas` only for the explicit value; anything else is the Câmara default. */
export const parseAcervoSource = (params: RawSearchParams): AcervoSource =>
firstValue(params[ACERVO_SOURCE_PARAM]) === ACERVO_SOURCE_ENVIADAS ? 'enviadas' : 'camara'
/**
* Parses the repeated `person` filter: trims, drops empty and oversized values
* (a label never exceeds the field limit, so an oversized token can only be a
* hand-typed miss) and dedupes case-insensitively keeping the first spelling.
*/
export const parseRecordingPeople = (raw: string | string[] | undefined): string[] => {
const people: string[] = []
const seen = new Set<string>()
for (const value of allParamValues(raw)) {
if (value.length > RECORDING_SPEAKER_LABEL_MAX_LENGTH) continue
const folded = foldSpeakerName(value)
if (seen.has(folded)) continue
seen.add(folded)
people.push(value)
}
return people
}
export const parseRecordingListParams = (params: RawSearchParams): RecordingListState => {
const rawPage = strictDecimalInteger(firstValue(params.page))
const rawQ = normalizedText(firstValue(params.q))
const q = rawQ && isContactSearchQueryReady(rawQ) ? rawQ : undefined
const people = parseRecordingPeople(params[ACERVO_PERSON_PARAM])
return {
source: 'enviadas',
page: rawPage ?? 1,
...(q ? { q } : {}),
...(people.length > 0 ? { people } : {}),
}
}
@ -56,10 +82,35 @@ const recordingListSearchParams = (
// `source=enviadas` is always serialized — it is what selects the source.
params.set(ACERVO_SOURCE_PARAM, ACERVO_SOURCE_ENVIADAS)
if (state.q) params.set('q', state.q)
for (const person of state.people ?? []) params.append(ACERVO_PERSON_PARAM, person)
if (page > 1) params.set('page', String(page))
return params
}
/** The URL of one "Pessoa" facet option with `person` toggled (page resets). */
export const toggleRecordingPerson = (
state: RecordingListState,
value: string,
): RecordingListState => {
const current = state.people ?? []
const folded = foldSpeakerName(value)
const selected = current.some((person) => foldSpeakerName(person) === folded)
const people = selected
? current.filter((person) => foldSpeakerName(person) !== folded)
: [...current, value]
return { ...state, page: 1, ...(people.length > 0 ? { people } : { people: undefined }) }
}
/** The URL that removes one selected "Pessoa" (chips with an X; page resets). */
export const removeRecordingPerson = (
state: RecordingListState,
value: string,
): RecordingListState => {
const folded = foldSpeakerName(value)
const people = (state.people ?? []).filter((person) => foldSpeakerName(person) !== folded)
return { ...state, page: 1, ...(people.length > 0 ? { people } : { people: undefined }) }
}
export const buildRecordingListHref = (state: RecordingListState, page: number): string =>
buildListHref(state, recordingListSearchParams, CAMPAIGN_COMMUNICATION_ACERVO, page)

View file

@ -2,6 +2,7 @@ import 'server-only'
import type { Payload } from 'payload'
import { foldSpeakerName } from '@/lib/recordingDiarization'
import { RECORDING_NOT_FOUND_MESSAGE } from '@/lib/schemas/recording'
import { normalizeForSearch } from '@/lib/speechSearch'
import type { CampaignUser } from '@/payload-types'
@ -33,29 +34,76 @@ const recordingListSelect = {
step: true,
recordedAt: true,
durationSeconds: true,
speakerNames: true,
} as const
const recordingDetailSelect = {
...recordingListSelect,
error: true,
speakerLabels: true,
speakerLabelsDropped: true,
} as const
const segmentSelect = {
startSeconds: true,
endSeconds: true,
text: true,
speakerKey: true,
} as const
const toSegmentRecord = (segment: {
startSeconds: number
endSeconds: number
text: string
speakerKey?: string | null
}): RecordingSegmentRecord => ({
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
text: segment.text,
speakerKey: segment.speakerKey ?? null,
})
/** C200 — the options the "Pessoa" facet offers. */
export type RecordingFilterOptions = {
people: string[]
}
/**
* C200 — distinct labels across the recordings the actor can read: the facet
* only ever lists what is reachable, and the derived `speakerNames` column is
* the single source. Small acervo now; if the list grows past a few thousand
* rows this becomes a `SELECT DISTINCT` (registered as a revisit trigger in the
* impl plan).
*/
export const loadRecordingFilterOptions = async (
payload: Payload,
user: CampaignUser,
): Promise<RecordingFilterOptions> => {
const result = await payload.find({
collection: 'recording',
depth: 0,
limit: 0,
pagination: false,
select: { speakerNames: true },
user,
overrideAccess: false,
})
const people: string[] = []
const seen = new Set<string>()
for (const recording of result.docs) {
for (const name of recording.speakerNames ?? []) {
const trimmed = name.trim()
const folded = foldSpeakerName(trimmed)
if (!trimmed || seen.has(folded)) continue
seen.add(folded)
people.push(trimmed)
}
}
return { people: people.sort((left, right) => left.localeCompare(right, 'pt-BR')) }
}
/**
* The excerpt segment of each listed recording: ONE matching segment per row
* (the search runs `LIKE` on the joined `recording.searchText`, so a page of
@ -120,6 +168,7 @@ const loadMatchedSegments = async (
export type RecordingsPageData = {
rows: RecordingListItemViewModel[]
state: RecordingListState
filterOptions: RecordingFilterOptions
redirectHref?: string
totalDocs: number
totalPages: number
@ -134,17 +183,20 @@ export const loadRecordingsPageData = async (
const canonicalUrl = resolveRecordingListUrl(rawSearchParams)
const state = canonicalUrl.state
const result = await payload.find({
collection: 'recording',
depth: 0,
limit: recordingPageSize,
page: state.page,
sort: '-createdAt',
where: buildRecordingListWhere(state),
select: recordingListSelect,
user,
overrideAccess: false,
})
const [result, filterOptions] = await Promise.all([
payload.find({
collection: 'recording',
depth: 0,
limit: recordingPageSize,
page: state.page,
sort: '-createdAt',
where: buildRecordingListWhere(state),
select: recordingListSelect,
user,
overrideAccess: false,
}),
loadRecordingFilterOptions(payload, user),
])
const resolvedUrl = resolveRecordingListUrl(rawSearchParams, result.totalPages)
const recordings = result.docs
@ -163,9 +215,11 @@ export const loadRecordingsPageData = async (
recording,
matchedSegments: segmentsByRecording.get(recording.id) ?? [],
query: state.q,
people: state.people,
}),
),
state: resolvedUrl.state,
filterOptions,
redirectHref: resolvedUrl.redirectHref ?? canonicalUrl.redirectHref,
totalDocs: result.totalDocs,
totalPages: result.totalPages,

View file

@ -1,13 +1,15 @@
/**
* C199 — view models of the uploaded recordings. Pure: the loader hands raw
* Payload rows and the view model decides labels, the matching excerpt and the
* links the list/detail render.
* Payload rows and the view model decides labels, the matching excerpt, the
* speaker groups and the links the list/detail render. C200 adds the anonymous
* cluster grouping ("Falante N") and the human labels the team typed.
*/
import {
recordingFailureMessage,
toRecordingViewModel,
type RecordingViewModel,
} from '@/lib/recording'
import { recordingSpeakerDefaultLabel, speakerNameMatches } from '@/lib/recordingDiarization'
import { formatSpeechClock } from '@/lib/speechClock'
import {
buildHighlightedExcerpt,
@ -22,6 +24,8 @@ export type RecordingSegmentRecord = {
startSeconds: number
endSeconds: number
text: string
/** C200 — anonymous cluster key; null/absent on recordings without diarization. */
speakerKey?: string | null
}
type RecordingBaseRecord = {
@ -33,15 +37,24 @@ type RecordingBaseRecord = {
durationSeconds?: number | null
}
export type RecordingListRecord = RecordingBaseRecord
export type RecordingListRecord = RecordingBaseRecord & {
/** C200 — derived distinct labels; drives the card chip under a "Pessoa" filter. */
speakerNames?: string[] | null
}
export type RecordingDetailRecord = RecordingBaseRecord & {
/** Raw internal cause of a failure; never handed to the client as-is. */
error?: string | null
/** C200 — human labels by cluster key. */
speakerLabels?: { speakerKey: string; label: string }[] | null
/** C200 — true when a reprocessing could not keep some label. */
speakerLabelsDropped?: boolean | null
}
export type RecordingListItemViewModel = RecordingViewModel & {
excerpt: SpeechHighlightedExcerpt | null
/** Selected "Pessoa" labels that this recording indeed contains. */
matchedPersons: string[]
/**
* Detail link that seeks the player to the matching segment when the search
* found one (`?t=`), carrying the term for transcript highlighting (`?q=`).
@ -56,10 +69,74 @@ export type RecordingDetailSegmentViewModel = {
parts: SpeechHighlightPart[]
}
/**
* C200 — one anonymous speaker cluster of the detail transcript: the keyed
* segments in order, the numbered default label and the human label when the
* team identified the group.
*/
export type RecordingSpeakerGroupViewModel = {
key: string
index: number
defaultLabel: string
label: string | null
isIdentified: boolean
segments: RecordingDetailSegmentViewModel[]
}
export type RecordingDetailViewModel = RecordingViewModel & {
/** Honest cause of a failure; null unless `status === 'failed'`. */
failureMessage: string | null
segments: RecordingDetailSegmentViewModel[]
/** Empty when the recording has no (or a partial) speaker grouping. */
speakerGroups: RecordingSpeakerGroupViewModel[]
/** C200 — surfaced by the imprecision banner when a label was dropped. */
speakerLabelsDropped: boolean
}
const toDetailSegment = (
segment: RecordingSegmentRecord,
query?: string,
): RecordingDetailSegmentViewModel => ({
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
startLabel: formatSpeechClock(segment.startSeconds),
parts: splitHighlightedParts(segment.text, query ?? ''),
})
/**
* Groups the transcript by cluster key. A recording only renders groups when
* EVERY segment is keyed: a half-diarized transcript would put an invented
* boundary next to a real one, so the partial case falls back to the plain
* list. Labels are looked up by key; an unlabeled group keeps "Falante N".
*/
const buildSpeakerGroups = (
segments: readonly RecordingSegmentRecord[],
labelsByKey: ReadonlyMap<string, string>,
query?: string,
): RecordingSpeakerGroupViewModel[] => {
if (segments.length === 0 || segments.some((segment) => !segment.speakerKey)) return []
const groups = new Map<string, RecordingDetailSegmentViewModel[]>()
for (const segment of segments) {
const key = segment.speakerKey
if (!key) continue
const group = groups.get(key)
const view = toDetailSegment(segment, query)
if (group) group.push(view)
else groups.set(key, [view])
}
return [...groups.entries()].map(([key, groupSegments], index) => {
const label = labelsByKey.get(key)?.trim() || null
return {
key,
index,
defaultLabel: recordingSpeakerDefaultLabel(index + 1),
label,
isIdentified: label !== null,
segments: groupSegments,
}
})
}
const withSeek = (href: string, segment: RecordingSegmentRecord | undefined, query?: string) => {
@ -80,28 +157,36 @@ export const toRecordingListItemViewModel = ({
recording,
matchedSegments,
query,
people,
}: {
recording: RecordingListRecord
/** Segments that matched the query (or the row's first one as fallback). */
matchedSegments: readonly RecordingSegmentRecord[]
query?: string
/** C200 — the selected "Pessoa" labels of the list state. */
people?: readonly string[]
}): RecordingListItemViewModel => {
const base = toRecordingViewModel(recording)
const isReady = base.status === 'ready'
const q = query?.trim()
const segment = isReady && q ? pickMatchingSegment(matchedSegments, q) : undefined
const excerpt = segment ? buildHighlightedExcerpt(segment.text, q ?? '') : null
const speakerNames = recording.speakerNames ?? []
return {
...base,
excerpt,
matchedPersons: (people ?? []).filter((person) =>
speakerNames.some((name) => speakerNameMatches(name, person)),
),
watchHref: withSeek(base.detailHref, segment, q),
}
}
/**
* Detail view of one recording: the base view plus the failure message and the
* clickable transcript segments (highlighted by the `?q=` of the search).
* Detail view of one recording: the base view plus the failure message, the
* clickable transcript segments (highlighted by the `?q=` of the search) and
* the speaker groups when the whole transcript is keyed.
*/
export const toRecordingDetailViewModel = ({
recording,
@ -114,6 +199,9 @@ export const toRecordingDetailViewModel = ({
}): RecordingDetailViewModel => {
const base = toRecordingViewModel(recording)
const isFailed = base.status === 'failed'
const labelsByKey = new Map(
(recording.speakerLabels ?? []).map((entry) => [entry.speakerKey, entry.label]),
)
return {
...base,
@ -123,11 +211,8 @@ export const toRecordingDetailViewModel = ({
step: base.step,
})
: null,
segments: segments.map((segment) => ({
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
startLabel: formatSpeechClock(segment.startSeconds),
parts: splitHighlightedParts(segment.text, query ?? ''),
})),
segments: segments.map((segment) => toDetailSegment(segment, query)),
speakerGroups: buildSpeakerGroups(segments, labelsByKey, query),
speakerLabelsDropped: recording.speakerLabelsDropped === true,
}
}

View file

@ -521,7 +521,18 @@ test.describe('communication vertical (C154/C162)', () => {
test.describe('uploaded recordings (C199)', () => {
const createRecording = async (
campaign: { fixtures: CampaignE2EOwnership },
input: { marker: string; status?: 'processing' | 'ready' | 'failed'; withSegments?: boolean },
input: {
marker: string
status?: 'processing' | 'ready' | 'failed'
withSegments?: boolean
segments?: {
startSeconds: number
endSeconds: number
text: string
speakerKey?: string | null
}[]
speakerLabels?: { speakerKey: string; label: string }[]
},
) => {
const bytes = Buffer.from(`recording-${input.marker}`)
const media = await campaign.fixtures.payload.create({
@ -541,11 +552,35 @@ test.describe('communication vertical (C154/C162)', () => {
status: input.status ?? 'ready',
recordedAt: '2026-09-01T00:00:00.000Z',
media: media.id,
...(input.speakerLabels ? { speakerLabels: input.speakerLabels } : {}),
},
depth: 0,
})
if (input.withSegments ?? true) {
if (input.segments) {
for (const [index, segment] of input.segments.entries()) {
await campaign.fixtures.payload.create({
collection: 'recordingSegment',
data: hookFilledCreateData<'recordingSegment'>({
recording: recording.id,
order: index + 1,
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
text: segment.text,
speakerKey: segment.speakerKey ?? null,
}),
depth: 0,
})
}
await campaign.fixtures.payload.update({
collection: 'recording',
id: recording.id,
data: {
searchText: normalizeForSearch(input.segments.map((segment) => segment.text).join(' ')),
},
depth: 0,
})
} else if (input.withSegments ?? true) {
await campaign.fixtures.payload.create({
collection: 'recordingSegment',
data: hookFilledCreateData<'recordingSegment'>({
@ -685,5 +720,155 @@ test.describe('communication vertical (C154/C162)', () => {
const denied = await advisorRequest.post(uploadUrl)
expect(denied.status()).toBe(403)
})
test('the grouped detail renders the speaker blocks, the warning and the identify action', async ({
campaign,
campaignRequest,
}) => {
const marker = campaign.fixtures.value('gravacaofalantes')
const { recording } = await createRecording(campaign, {
marker,
segments: [
{
startSeconds: 12,
endSeconds: 20,
text: `A plenária discutiu a merenda ${marker}.`,
speakerKey: 'speaker-1',
},
{
startSeconds: 21,
endSeconds: 30,
text: 'O relator respondeu em seguida.',
speakerKey: 'speaker-2',
},
],
speakerLabels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }],
})
const user = await campaign.fixtures.createCampaignUser('communicator')
const request = await campaignRequest(user, user.password)
const response = await request.get(
`/campanha/comunicacao/acervo/gravacoes/${recording.id}?t=12&q=${marker}`,
)
expect(response.status()).toBe(200)
const html = rendered(await response.text())
expect(html).toContain('Transcrição por falante')
expect(html).toContain('Dep. Jorge Solla')
expect(html).toContain('Identificação feita pela equipe')
expect(html).toContain('Falante 2')
expect(html).toContain('Agrupamento sem identificação')
expect(html).toContain('Editar identificação')
expect(html).toContain('Identificar falante')
expect(html).toContain('A separação por falante pode estar imprecisa')
expect(html).toContain('data-start-seconds="12"')
expect(html).toContain('data-start-seconds="21"')
expect(html).toContain('<mark')
})
test('the "Pessoa" facet filters the list and the card shows the chip', async ({
campaign,
campaignRequest,
}) => {
const marker = campaign.fixtures.value('gravacaopessoa')
const { recording } = await createRecording(campaign, {
marker,
segments: [
{
startSeconds: 12,
endSeconds: 20,
text: `A plenária discutiu a merenda ${marker}.`,
speakerKey: 'speaker-1',
},
],
speakerLabels: [{ speakerKey: 'speaker-1', label: `Dep. ${marker}` }],
})
const { recording: other } = await createRecording(campaign, {
marker: `${marker}outra`,
withSegments: false,
})
const user = await campaign.fixtures.createCampaignUser('communicator')
const request = await campaignRequest(user, user.password)
const response = await request.get(
`/campanha/comunicacao/acervo?source=enviadas&person=${encodeURIComponent(`dep. ${marker}`)}`,
)
expect(response.status()).toBe(200)
const html = rendered(await response.text())
expect(html).toContain(`Gravação ${marker}`)
expect(html).not.toContain(`Gravação ${marker}outra`)
expect(html).toContain('Pessoa:')
expect(html).toContain('aparece nesta gravação')
expect(html).toContain(`/campanha/comunicacao/acervo/gravacoes/${recording.id}`)
expect(html).not.toContain(`/campanha/comunicacao/acervo/gravacoes/${other.id}`)
})
test('labels a cluster through the JSON route and denies the non-acervo roles', async ({
campaign,
campaignRequest,
}) => {
const marker = campaign.fixtures.value('gravacaorotular')
const { recording } = await createRecording(campaign, {
marker,
segments: [
{
startSeconds: 12,
endSeconds: 20,
text: 'Fala única do agrupamento.',
speakerKey: 'speaker-1',
},
],
})
const endpoint = `/campanha/comunicacao/acervo/gravacoes/${recording.id}/falantes`
const user = await campaign.fixtures.createCampaignUser('communicator')
const request = await campaignRequest(user, user.password)
const response = await request.post(endpoint, {
data: { recordingId: recording.id, speakerKey: 'speaker-1', label: `Dep. ${marker}` },
})
expect(response.status()).toBe(200)
expect((await response.json()).status).toBe('success')
const updated = await campaign.fixtures.payload.findByID({
collection: 'recording',
id: recording.id,
depth: 0,
})
expect(updated.speakerNames).toEqual([`Dep. ${marker}`])
const unknown = await request.post(endpoint, {
data: { recordingId: recording.id, speakerKey: 'speaker-9', label: 'Fantasma' },
})
expect(unknown.status()).toBe(400)
const advisor = await campaign.fixtures.createCampaignUser('advisor')
const advisorRequest = await campaignRequest(advisor, advisor.password)
const denied = await advisorRequest.post(endpoint, {
data: { recordingId: recording.id, speakerKey: 'speaker-1', label: 'Negado' },
})
// The JSON mutation wrapper maps a safe domain message to 400 (same
// envelope as the other `/campanha/**` mutation routes).
expect(denied.status()).toBe(400)
})
test('a recording without diarization keeps the plain transcript', async ({
campaign,
campaignRequest,
}) => {
const marker = campaign.fixtures.value('gravacaoSemFalantes')
const { recording } = await createRecording(campaign, { marker })
const user = await campaign.fixtures.createCampaignUser('communicator')
const request = await campaignRequest(user, user.password)
const response = await request.get(
`/campanha/comunicacao/acervo/gravacoes/${recording.id}?t=12&q=${marker}`,
)
const html = rendered(await response.text())
expect(html).toContain('data-start-seconds="12"')
expect(html).not.toContain('Transcrição por falante')
expect(html).not.toContain('A separação por falante pode estar imprecisa')
})
})
})

View file

@ -29,9 +29,14 @@ import { GET } from '@/app/(campaign)/campanha/(app)/comunicacao/acervo/gravacoe
import {
deleteRecordingForActor,
getRecordingStatusesForActor,
labelRecordingSpeakerForActor,
retryRecordingForActor,
} from '@/app/(campaign)/campanha/actions/recording'
import { RECORDING_MAX_BYTES } from '@/lib/recording'
import {
RECORDING_FORBIDDEN_MESSAGE,
RECORDING_SPEAKER_UNKNOWN_MESSAGE,
} from '@/lib/schemas/recording'
import { normalizeForSearch } from '@/lib/speechSearch'
import type { CampaignUser, RecordingMedia } from '@/payload-types'
import config from '@/payload.config'
@ -41,7 +46,10 @@ import {
RECORDING_STALE_MS,
runRecordingJob,
} from '@/utilities/recordings/recordingJob'
import { loadRecordingsPageData } from '@/utilities/recordings/recordingPageData'
import {
loadRecordingFilterOptions,
loadRecordingsPageData,
} from '@/utilities/recordings/recordingPageData'
import { receiveRecordingUpload } from '@/utilities/recordings/recordingUpload'
import { installCampaignFixtures } from '../helpers/campaignFixtures'
@ -90,11 +98,18 @@ const createRecording = async ({
status = 'ready',
title = 'Plenária da Comissão',
segments = [],
speakerLabels,
withMedia = true,
}: {
status?: 'uploading' | 'processing' | 'ready' | 'failed'
title?: string
segments?: { startSeconds: number; endSeconds: number; text: string }[]
segments?: {
startSeconds: number
endSeconds: number
text: string
speakerKey?: string | null
}[]
speakerLabels?: { speakerKey: string; label: string }[]
withMedia?: boolean
} = {}) => {
const media = withMedia ? await createMedia(`${title}.mp4`) : null
@ -104,6 +119,7 @@ const createRecording = async ({
title,
status,
...(media ? { media: media.id } : {}),
...(speakerLabels ? { speakerLabels } : {}),
},
overrideAccess: true,
})
@ -118,6 +134,7 @@ const createRecording = async ({
startSeconds: segment.startSeconds,
endSeconds: segment.endSeconds,
text: segment.text,
speakerKey: segment.speakerKey ?? null,
}),
overrideAccess: true,
})
@ -555,4 +572,197 @@ describe('uploaded recordings (C199)', () => {
expect(unfiltered.rows.length).toBeGreaterThanOrEqual(2)
expect(unfiltered.rows.every((row) => row.excerpt === null)).toBe(true)
})
it('filters recordings by a human label and lists the facet options', async () => {
const marker = `solla${Date.now().toString(36)}`
const { recording } = await createRecording({
title: `Plenária com vozes ${marker}`,
segments: [
{ startSeconds: 0, endSeconds: 5, text: 'Primeira fala.', speakerKey: 'speaker-1' },
{ startSeconds: 5, endSeconds: 10, text: 'Segunda fala.', speakerKey: 'speaker-2' },
],
speakerLabels: [{ speakerKey: 'speaker-1', label: `Dep. ${marker}` }],
})
const { recording: other } = await createRecording({
title: `Sem falantes ${marker}`,
segments: [],
})
// The facet matches the same way the dialog chip does: case-insensitive
// containment over the derived names, against the real join table.
const filtered = await loadRecordingsPageData(payload, communicator, {
source: 'enviadas',
person: [`dep. ${marker}`],
})
expect(filtered.rows.map((row) => row.id)).toContain(recording.id)
expect(filtered.rows.map((row) => row.id)).not.toContain(other.id)
expect(filtered.rows.find((row) => row.id === recording.id)?.matchedPersons).toEqual([
`dep. ${marker}`,
])
const options = await loadRecordingFilterOptions(payload, communicator)
expect(options.people).toContain(`Dep. ${marker}`)
})
it('labels one cluster through the action and rebuilds the facet names', async () => {
const marker = `falante${Date.now().toString(36)}`
const { recording } = await createRecording({
title: `Debate ${marker}`,
segments: [{ startSeconds: 0, endSeconds: 5, text: 'Fala única.', speakerKey: 'speaker-1' }],
})
getCampaignUserMock.mockResolvedValue(communicator)
await labelRecordingSpeakerForActor({
recordingId: recording.id,
speakerKey: 'speaker-1',
label: `Dep. ${marker}`,
})
const updated = await payload.findByID({
collection: 'recording',
id: recording.id,
depth: 0,
overrideAccess: true,
})
expect(updated.speakerLabels).toEqual([
expect.objectContaining({ speakerKey: 'speaker-1', label: `Dep. ${marker}` }),
])
expect(updated.speakerNames).toEqual([`Dep. ${marker}`])
for (const denied of [advisor, leader]) {
getCampaignUserMock.mockResolvedValue(denied)
await expect(
labelRecordingSpeakerForActor({
recordingId: recording.id,
speakerKey: 'speaker-1',
label: 'Quem quer que seja',
}),
).rejects.toThrow(RECORDING_FORBIDDEN_MESSAGE)
}
getCampaignUserMock.mockResolvedValue(communicator)
await expect(
labelRecordingSpeakerForActor({
recordingId: recording.id,
speakerKey: 'speaker-9',
label: 'Agrupamento fantasma',
}),
).rejects.toThrow(RECORDING_SPEAKER_UNKNOWN_MESSAGE)
})
it('groups the transcript with the injected diarizer and keeps a reconcilable label', async () => {
const { recording } = await createRecording({
status: 'processing',
title: 'Plenária com falantes',
segments: [
{ startSeconds: 1, endSeconds: 4, text: 'Fala anterior A', speakerKey: 'speaker-1' },
{ startSeconds: 5, endSeconds: 9, text: 'Fala anterior B', speakerKey: 'speaker-2' },
],
speakerLabels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }],
})
const restore = withEnv('FFMPEG_PATH', FAKE_FFMPEG)
try {
await runRecordingJob(
payload,
recording.id,
async () => ({
ok: true,
segments: [
{ start: 1, end: 4, text: 'Bom dia a todas' },
{ start: 5, end: 9, text: 'e a todos' },
],
durationSeconds: 1_200,
}),
async (file, options) => {
expect(file).toBeInstanceOf(Blob)
await options?.onProgress?.()
return {
ok: true,
turns: [
{ speaker: 'A', startSeconds: 0, endSeconds: 4.5 },
{ speaker: 'B', startSeconds: 4.5, endSeconds: 10 },
],
}
},
)
} finally {
restore()
}
const updated = await payload.findByID({
collection: 'recording',
id: recording.id,
depth: 0,
overrideAccess: true,
})
expect(updated.status).toBe('ready')
expect(updated.speakerLabels).toEqual([
expect.objectContaining({ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }),
])
expect(updated.speakerLabelsDropped).toBe(false)
const segments = await payload.find({
collection: 'recordingSegment',
where: { recording: { equals: recording.id } },
depth: 0,
limit: 0,
pagination: false,
sort: 'order',
overrideAccess: true,
})
expect(segments.docs).toHaveLength(4)
expect(segments.docs.map((segment) => segment.speakerKey)).toEqual([
'speaker-1',
'speaker-2',
'speaker-2',
'speaker-2',
])
})
it('keeps the plain transcript and warns when the diarizer fails with labels present', async () => {
const { recording } = await createRecording({
status: 'processing',
title: 'Debate sem provedor',
segments: [{ startSeconds: 1, endSeconds: 4, text: 'Fala', speakerKey: 'speaker-1' }],
speakerLabels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }],
})
const restore = withEnv('FFMPEG_PATH', FAKE_FFMPEG)
try {
await runRecordingJob(
payload,
recording.id,
async () => ({
ok: true,
segments: [{ start: 1, end: 4, text: 'Fala' }],
durationSeconds: 600,
}),
async () => ({ ok: false, error: 'provider down', status: 502 }),
)
} finally {
restore()
}
const updated = await payload.findByID({
collection: 'recording',
id: recording.id,
depth: 0,
overrideAccess: true,
})
expect(updated.status).toBe('ready')
expect(updated.speakerLabels ?? []).toEqual([])
expect(updated.speakerLabelsDropped).toBe(true)
const segments = await payload.find({
collection: 'recordingSegment',
where: { recording: { equals: recording.id } },
depth: 0,
limit: 0,
pagination: false,
sort: 'order',
overrideAccess: true,
})
expect(segments.docs.every((segment) => !segment.speakerKey)).toBe(true)
})
})

View file

@ -0,0 +1,134 @@
// @vitest-environment node
import { afterEach, describe, expect, it, vi } from 'vitest'
import {
ASSEMBLYAI_TRANSCRIPT_URL,
ASSEMBLYAI_UPLOAD_URL,
assemblyAiDiarize,
} from '@/utilities/ai/assemblyAiDiarize'
type FetchCall = { url: string; init: RequestInit | undefined }
const jsonResponse = (body: unknown, status = 200): Response =>
new Response(JSON.stringify(body), { status })
const audioBlob = (): Blob => new Blob([new Uint8Array([1, 2, 3])], { type: 'audio/mpeg' })
const installFetch = (
handler: (url: string, init: RequestInit | undefined, call: number) => Response,
): FetchCall[] => {
const calls: FetchCall[] = []
vi.stubGlobal(
'fetch',
vi.fn(async (input: RequestInfo | URL, init?: RequestInit) => {
const url = String(input)
calls.push({ url, init })
return handler(url, init, calls.length)
}),
)
return calls
}
afterEach(() => {
vi.unstubAllGlobals()
vi.unstubAllEnvs()
})
describe('AssemblyAI diarization provider (C200)', () => {
it('uploads, submits with speaker labels and normalizes ms to seconds', async () => {
vi.stubEnv('ASSEMBLYAI_API_KEY', 'test-key')
const calls = installFetch((url, init) => {
if (url === ASSEMBLYAI_UPLOAD_URL)
return jsonResponse({ upload_url: 'https://cdn/audio.mp3' })
if (url === ASSEMBLYAI_TRANSCRIPT_URL && init?.method === 'POST') {
return jsonResponse({ id: 'tr_1', status: 'queued' })
}
return jsonResponse({
status: 'completed',
utterances: [
{ speaker: 'A', start: 1000, end: 2500 },
{ speaker: 'B', start: 2500, end: 4000 },
{ speaker: 'C', start: 'x', end: 10 },
],
})
})
const progress = vi.fn()
const result = await assemblyAiDiarize(audioBlob(), {
pollIntervalMs: 1,
onProgress: progress,
})
expect(result).toEqual({
ok: true,
turns: [
{ speaker: 'A', startSeconds: 1, endSeconds: 2.5 },
{ speaker: 'B', startSeconds: 2.5, endSeconds: 4 },
],
})
const create = calls.find(
(call) => call.url === ASSEMBLYAI_TRANSCRIPT_URL && call.init?.method === 'POST',
)
expect(create?.init?.body).toContain('"speaker_labels":true')
expect(create?.init?.body).toContain('"speech_models":["universal-2"]')
expect(create?.init?.body).toContain('"language_code":"pt"')
expect(new Headers(calls[0]?.init?.headers).get('authorization')).toBe('test-key')
expect(calls.some((call) => call.init?.method === 'DELETE')).toBe(true)
expect(progress).toHaveBeenCalled()
})
it('answers 503 without a key and never calls the provider', async () => {
vi.stubEnv('ASSEMBLYAI_API_KEY', '')
const calls = installFetch(() => jsonResponse({}))
expect(await assemblyAiDiarize(audioBlob())).toEqual({
ok: false,
error: 'Separação de falantes está indisponível no momento.',
status: 503,
})
expect(calls).toHaveLength(0)
})
it('degrades to an error when the upload fails', async () => {
vi.stubEnv('ASSEMBLYAI_API_KEY', 'test-key')
const calls = installFetch(() => jsonResponse({}, 500))
const result = await assemblyAiDiarize(audioBlob())
expect(result.ok).toBe(false)
expect(calls).toHaveLength(1)
})
it('degrades to an error and deletes the remote transcript on provider error', async () => {
vi.stubEnv('ASSEMBLYAI_API_KEY', 'test-key')
const calls = installFetch((url, init) => {
if (url === ASSEMBLYAI_UPLOAD_URL)
return jsonResponse({ upload_url: 'https://cdn/audio.mp3' })
if (url === ASSEMBLYAI_TRANSCRIPT_URL && init?.method === 'POST') {
return jsonResponse({ id: 'tr_9', status: 'queued' })
}
if (init?.method === 'DELETE') return jsonResponse({}, 200)
return jsonResponse({ status: 'error', error: 'provider exploded' })
})
expect(await assemblyAiDiarize(audioBlob(), { pollIntervalMs: 1 })).toMatchObject({ ok: false })
expect(calls.some((call) => call.init?.method === 'DELETE')).toBe(true)
})
it('accepts a completed transcript without utterances (honest empty turns)', async () => {
vi.stubEnv('ASSEMBLYAI_API_KEY', 'test-key')
installFetch((url, init) => {
if (url === ASSEMBLYAI_UPLOAD_URL)
return jsonResponse({ upload_url: 'https://cdn/audio.mp3' })
if (url === ASSEMBLYAI_TRANSCRIPT_URL && init?.method === 'POST') {
return jsonResponse({ id: 'tr_2', status: 'queued' })
}
return jsonResponse({ status: 'completed', utterances: [] })
})
expect(await assemblyAiDiarize(audioBlob(), { pollIntervalMs: 1 })).toEqual({
ok: true,
turns: [],
})
})
})

View file

@ -21,6 +21,7 @@ import {
} from '@/lib/recording'
import {
buildRecordingAudioFfmpegArgs,
buildRecordingFullAudioFfmpegArgs,
mergeChunkTranscriptions,
RECORDING_AUDIO_CHUNK_SECONDS,
recordingSearchText,
@ -180,4 +181,17 @@ describe('recording transcription (C199)', () => {
expect(mergeChunkTranscriptions([])).toEqual([])
expect(recordingSearchText([])).toBe('')
})
it('builds the single-file extraction for diarization (C200)', () => {
const args = buildRecordingFullAudioFfmpegArgs({
inputPath: '/tmp/in.mkv',
outputPath: '/tmp/full.mp3',
})
expect(args).toContain('-vn')
expect(args).toContain('libmp3lame')
// One output, no segmentation: speaker numbering needs the whole audio.
expect(args).not.toContain('-f')
expect(args).not.toContain('-segment_time')
expect(args.at(-1)).toBe('/tmp/full.mp3')
})
})

View file

@ -0,0 +1,177 @@
// @vitest-environment node
import { describe, expect, it } from 'vitest'
import {
assignSpeakerKeys,
reconcileSpeakerLabels,
RECORDING_DIARIZATION_MAX_SECONDS,
recordingSpeakerDefaultLabel,
speakerKeysInOrder,
speakerNameMatches,
speakerNamesFromLabels,
type DiarizedTurn,
} from '@/lib/recordingDiarization'
const segment = (startSeconds: number, endSeconds: number) => ({
startSeconds,
endSeconds,
text: `trecho ${startSeconds}`,
})
const turn = (speaker: string, startSeconds: number, endSeconds: number): DiarizedTurn => ({
speaker,
startSeconds,
endSeconds,
})
describe('recording speaker grouping (C200)', () => {
it('assigns each segment to the turn with the largest overlap', () => {
const segments = [segment(0, 10), segment(10, 20), segment(20, 30)]
const turns = [turn('A', 0, 11), turn('B', 12, 30)]
const keyed = assignSpeakerKeys(segments, turns)
expect(keyed.map((entry) => entry.speakerKey)).toEqual(['speaker-1', 'speaker-2', 'speaker-2'])
// The original fields survive the assignment.
expect(keyed[1]).toMatchObject({ startSeconds: 10, endSeconds: 20, text: 'trecho 10' })
})
it('breaks an overlap tie by the closest midpoint', () => {
const segments = [segment(0, 10)]
// Both turns overlap the same 5 s; B's midpoint is closer to the segment's.
const turns = [turn('B', 5, 20), turn('A', 0, 5)]
expect(assignSpeakerKeys(segments, turns)[0]?.speakerKey).toBe('speaker-1')
expect(
assignSpeakerKeys([segment(0, 10)], [turn('A', 0, 5), turn('B', 5, 20)])[0],
).toMatchObject({
speakerKey: 'speaker-1',
})
})
it('falls back to the nearest turn when there is no overlap at all', () => {
const segments = [segment(0, 10), segment(30, 40)]
const turns = [turn('A', 12, 20), turn('B', 45, 60)]
expect(assignSpeakerKeys(segments, turns).map((entry) => entry.speakerKey)).toEqual([
'speaker-1',
'speaker-2',
])
})
it('numbers clusters globally by first appearance, not by provider name', () => {
const segments = [segment(0, 5), segment(5, 10), segment(10, 15)]
const turns = [turn('spk_9', 10, 15), turn('spk_2', 0, 5), turn('spk_9', 0, 10)]
expect(assignSpeakerKeys(segments, turns).map((entry) => entry.speakerKey)).toEqual([
'speaker-1',
'speaker-2',
'speaker-2',
])
})
it('keeps every key null without a usable turn and ignores invalid turns', () => {
expect(assignSpeakerKeys([segment(0, 5)], []).map((entry) => entry.speakerKey)).toEqual([null])
expect(
assignSpeakerKeys([segment(0, 5)], [turn('', 0, 5), turn('A', Number.NaN, 5)]).map(
(entry) => entry.speakerKey,
),
).toEqual([null])
})
it('lists the cluster keys in first-appearance order', () => {
expect(
speakerKeysInOrder([
{ speakerKey: 'speaker-2' },
{ speakerKey: null },
{ speakerKey: 'speaker-1' },
{ speakerKey: 'speaker-2' },
]),
).toEqual(['speaker-2', 'speaker-1'])
})
it('reconciles labels 1:1 by time overlap and reports what was dropped', () => {
const previousSegments = [{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 10 }]
const nextSegments = [{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 10 }]
expect(
reconcileSpeakerLabels({
previousLabels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }],
previousSegments,
nextSegments,
}),
).toEqual({ labels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }], dropped: false })
expect(
reconcileSpeakerLabels({
previousLabels: [{ speakerKey: 'speaker-1', label: 'Dep. Jorge Solla' }],
previousSegments,
nextSegments: [{ speakerKey: 'speaker-1', startSeconds: 40, endSeconds: 50 }],
}),
).toEqual({ labels: [], dropped: true })
expect(
reconcileSpeakerLabels({
previousLabels: [],
previousSegments,
nextSegments,
}),
).toEqual({ labels: [], dropped: false })
})
it('never shares one new cluster between two old labels (merge)', () => {
// Two old clusters merged into one new cluster: keeping either label would
// name the other person's speech, so both are dropped.
const result = reconcileSpeakerLabels({
previousLabels: [
{ speakerKey: 'speaker-1', label: 'Solla' },
{ speakerKey: 'speaker-2', label: 'Outra pessoa' },
],
previousSegments: [
{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 5 },
{ speakerKey: 'speaker-2', startSeconds: 5, endSeconds: 10 },
],
nextSegments: [{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 10 }],
})
expect(result).toEqual({ labels: [], dropped: true })
})
it('never duplicates one label across two new clusters (split)', () => {
// One old cluster split into two new ones: the label cannot name both.
const result = reconcileSpeakerLabels({
previousLabels: [{ speakerKey: 'speaker-1', label: 'Solla' }],
previousSegments: [{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 10 }],
nextSegments: [
{ speakerKey: 'speaker-1', startSeconds: 0, endSeconds: 5 },
{ speakerKey: 'speaker-2', startSeconds: 5, endSeconds: 10 },
],
})
expect(result).toEqual({ labels: [], dropped: true })
})
it('derives distinct speaker names case-insensitively, in order', () => {
expect(
speakerNamesFromLabels([
{ speakerKey: 'speaker-1', label: ' Dep. Jorge Solla ' },
{ speakerKey: 'speaker-2', label: 'dep. jorge solla' },
{ speakerKey: 'speaker-3', label: 'Presidente' },
{ speakerKey: 'speaker-4', label: ' ' },
]),
).toEqual(['Dep. Jorge Solla', 'Presidente'])
expect(speakerNamesFromLabels(null)).toEqual([])
})
it('matches a facet term the same way the ILIKE query does', () => {
expect(speakerNameMatches('Dep. Jorge Solla', 'solla')).toBe(true)
expect(speakerNameMatches('Dep. Jorge Solla', 'Dep. Jorge')).toBe(true)
expect(speakerNameMatches('Presidente', 'solla')).toBe(false)
})
it('names the default label with a number and pins the provider ceiling', () => {
expect(recordingSpeakerDefaultLabel(1)).toBe('Falante 1')
expect(recordingSpeakerDefaultLabel(12)).toBe('Falante 12')
expect(RECORDING_DIARIZATION_MAX_SECONDS).toBe(10 * 60 * 60)
})
})

View file

@ -0,0 +1,44 @@
// @vitest-environment node
import { describe, expect, it } from 'vitest'
import { buildRecordingListWhere } from '@/utilities/recordings/recordingListFilters'
describe('recording list filters (C199/C200)', () => {
it('returns an empty where without search or facet', () => {
expect(buildRecordingListWhere({ source: 'enviadas', page: 1 })).toEqual({})
})
it('searches the normalized transcript text', () => {
expect(buildRecordingListWhere({ source: 'enviadas', page: 1, q: 'Comissão' })).toEqual({
searchText: { like: 'comissao' },
})
})
it('filters one person through the derived names', () => {
expect(
buildRecordingListWhere({ source: 'enviadas', page: 1, people: ['Dep. Jorge Solla'] }),
).toEqual({ speakerNames: { contains: 'Dep. Jorge Solla' } })
})
it('ORs the selected people and ANDs them with the search', () => {
expect(
buildRecordingListWhere({
source: 'enviadas',
page: 1,
q: 'merenda',
people: ['Solla', 'Presidente'],
}),
).toEqual({
and: [
{ searchText: { like: 'merenda' } },
{
or: [
{ speakerNames: { contains: 'Solla' } },
{ speakerNames: { contains: 'Presidente' } },
],
},
],
})
})
})

View file

@ -7,7 +7,10 @@ import {
buildRecordingListHref,
parseAcervoSource,
parseRecordingListParams,
parseRecordingPeople,
removeRecordingPerson,
resolveRecordingListUrl,
toggleRecordingPerson,
} from '@/utilities/recordings/recordingListUrl'
describe('recording list URL (C199)', () => {
@ -68,4 +71,63 @@ describe('recording list URL (C199)', () => {
expect(buildAcervoSourceHref('enviadas')).toBe('/campanha/comunicacao/acervo?source=enviadas')
expect(buildAcervoSourceHref('camara')).toBe('/campanha/comunicacao/acervo')
})
it('parses repeated `person` values: trim, oversized drop, case-insensitive dedupe', () => {
expect(parseRecordingPeople([' Solla ', 'solla', 'Presidente'])).toEqual([
'Solla',
'Presidente',
])
expect(parseRecordingPeople(undefined)).toEqual([])
expect(parseRecordingPeople('')).toEqual([])
expect(parseRecordingPeople('x'.repeat(121))).toEqual([])
})
it('canonicalizes and serializes the `person` facet with the query', () => {
const parsed = parseRecordingListParams({
source: 'enviadas',
q: 'merenda',
person: ['Solla', 'Presidente'],
})
expect(parsed.people).toEqual(['Solla', 'Presidente'])
const canonical = resolveRecordingListUrl({
source: 'enviadas',
q: 'merenda',
person: ['Solla', 'Presidente'],
})
expect(canonical.href).toBe(
'/campanha/comunicacao/acervo?source=enviadas&q=merenda&person=Solla&person=Presidente',
)
expect(canonical.redirectHref).toBeUndefined()
})
it('drops a duplicated facet value on canonicalization', () => {
const canonical = resolveRecordingListUrl({
source: 'enviadas',
person: ['Solla', 'solla', 'x'.repeat(200)],
})
expect(canonical.href).toBe('/campanha/comunicacao/acervo?source=enviadas&person=Solla')
expect(canonical.redirectHref).toBe('/campanha/comunicacao/acervo?source=enviadas&person=Solla')
})
it('toggles and removes one person, resetting to the first page', () => {
const base = { source: 'enviadas' as const, page: 3, q: 'merenda' }
const added = toggleRecordingPerson(base, 'Solla')
expect(added).toEqual({ ...base, page: 1, people: ['Solla'] })
const removed = toggleRecordingPerson({ ...base, people: ['Solla'] }, 'solla')
expect(removed).toEqual({ ...base, page: 1, people: undefined })
expect(removeRecordingPerson({ ...base, people: ['Solla', 'Presidente'] }, 'Solla')).toEqual({
...base,
page: 1,
people: ['Presidente'],
})
expect(removeRecordingPerson({ ...base, people: ['Solla'] }, 'Solla')).toEqual({
...base,
page: 1,
people: undefined,
})
})
})