import 'server-only' import { normalizeRecordingTranscription, type RecordingChunkSegment, } from '@/lib/recordingTranscription' /** * Transcribes a recorded voice clip to text via Deep Infra's OpenAI-compatible * transcription endpoint (B173). Deep Infra serves Whisper large-v3 at * $0.00045/audio minute — the cheapest speech-to-text around — so it stays * here even though the Sollinha LLM talks to api.deepseek.com directly * (`DEEPSEEK_API_KEY`); this path uses its own `DEEPINFRA_API_KEY`. * * The audio is forwarded as a multipart `file` and never persisted anywhere — * it lives only in this function's memory for the duration of the provider * call (product requirement: B173 processes and discards the recording). * * C199 — `deepInfraTranscribeSegments` is the timestamped sibling used by the * recording transcription job: same provider and key, `verbose_json` with * segment timestamps (the shape the Câmara CLI already validates). */ export const DEEPINFRA_TRANSCRIBE_URL = 'https://api.deepinfra.com/v1/openai/audio/transcriptions' const DEEPINFRA_WHISPER_MODEL = 'openai/whisper-large-v3' const DEEPINFRA_LANGUAGE = 'pt' /** * C215 — provider price (~US$ per audio minute, Whisper large-v3 at Deep * Infra) used by the ingestion report to state the ASR cost honestly. The * Câmara CLI keeps its own copy; this is the app-side owner. */ export const DEEPINFRA_TRANSCRIBE_COST_PER_MINUTE_USD = 0.00045 export type TranscribeAudioResult = | { ok: true; text: string } | { ok: false; error: string; status: number } const TRANSCRIBE_ERROR = 'Não foi possível transcrever o áudio. Tente novamente.' /** * Sends the clip to Deep Infra Whisper large-v3 and returns the transcript. * Never throws: every failure becomes `{ ok: false }` with a pt-BR message the * route can return verbatim, so a provider outage or a missing key degrades to * "sem voz" instead of breaking the text chat. */ export const deepInfraTranscribe = async (file: Blob): Promise => { const apiKey = process.env.DEEPINFRA_API_KEY if (!apiKey) { return { ok: false, error: 'Transcrição de voz está indisponível no momento.', status: 503 } } const form = new FormData() form.append('file', file, 'voice.webm') form.append('model', DEEPINFRA_WHISPER_MODEL) form.append('language', DEEPINFRA_LANGUAGE) try { const response = await fetch(DEEPINFRA_TRANSCRIBE_URL, { method: 'POST', headers: { Authorization: `Bearer ${apiKey}` }, body: form, }) if (!response.ok) { return { ok: false, error: TRANSCRIBE_ERROR, status: 502, } } const data = (await response.json()) as { text?: unknown } const text = typeof data.text === 'string' ? data.text.trim() : '' if (!text) { return { ok: false, error: TRANSCRIBE_ERROR, status: 502, } } return { ok: true, text } } catch { return { ok: false, error: TRANSCRIBE_ERROR, status: 502, } } } export type TranscribeSegmentsResult = | { ok: true; segments: RecordingChunkSegment[]; durationSeconds: number | null } | { ok: false; error: string; status: number } const SEGMENTS_TIMEOUT_MS = 600_000 /** * C199 — sends one audio chunk to Deep Infra Whisper with segment timestamps * and returns the normalized segments (relative to the chunk). Never throws: * the job maps `{ ok: false }` to the recording's `failed` state with the * pt-BR message. The caller owns the offset merge and the chunk size. */ export const deepInfraTranscribeSegments = async ( file: Blob, { filename = 'chunk.mp3', timeoutMs = SEGMENTS_TIMEOUT_MS, }: { filename?: string; timeoutMs?: number } = {}, ): Promise => { const apiKey = process.env.DEEPINFRA_API_KEY if (!apiKey) { return { ok: false, error: 'Transcrição de áudio está indisponível no momento.', status: 503 } } const form = new FormData() form.append('file', file, filename) form.append('model', DEEPINFRA_WHISPER_MODEL) form.append('language', DEEPINFRA_LANGUAGE) form.append('response_format', 'verbose_json') form.append('timestamp_granularities[]', 'segment') try { const response = await fetch(DEEPINFRA_TRANSCRIBE_URL, { method: 'POST', headers: { Authorization: `Bearer ${apiKey}` }, body: form, signal: AbortSignal.timeout(timeoutMs), }) if (!response.ok) { return { ok: false, error: TRANSCRIBE_ERROR, status: 502 } } const normalized = normalizeRecordingTranscription(await response.json()) if (normalized.segments.length === 0) { return { ok: false, error: TRANSCRIBE_ERROR, status: 502 } } return { ok: true, ...normalized } } catch { return { ok: false, error: TRANSCRIBE_ERROR, status: 502 } } }