import logging import httpx from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type from app.config import settings logger = logging.getLogger(__name__) MAX_AUDIO_SIZE_BYTES = 10 * 1024 * 1024 # 10 MB @retry( stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=1, max=5), retry=retry_if_exception_type(httpx.HTTPStatusError), reraise=True, ) async def transcribe_audio(audio_bytes: bytes, language: str = "gu-In") -> dict: """ Send audio to Sarvam AI Saaras v3 API and return the transcript. Args: audio_bytes: Raw audio bytes (WebM/Opus from MediaRecorder). language: Language hint (e.g. 'gu-In', 'hi-In', 'en-In'). Returns: dict with 'transcript' and 'language_detected' keys. """ if len(audio_bytes) > MAX_AUDIO_SIZE_BYTES: raise ValueError(f"Audio file too large: {len(audio_bytes)} bytes (max {MAX_AUDIO_SIZE_BYTES})") async with httpx.AsyncClient(timeout=30.0) as client: response = await client.post( settings.sarvam_api_url, headers={"api-subscription-key": settings.sarvam_api_key}, files={"file": ("audio.webm", audio_bytes, "audio/webm")}, data={ "model": "saaras:v3", "mode": "transcribe", "language_code": language, }, ) if response.status_code == 400: logger.warning("Sarvam API 400: %s", response.text) raise httpx.HTTPStatusError( f"Sarvam API error: {response.text}", request=response.request, response=response, ) response.raise_for_status() result = response.json() transcript = result.get("transcript", "") lang_detected = result.get("language_code", language) logger.info("Sarvam STT: '%s...' (lang=%s)", transcript[:60], lang_detected) return { "transcript": transcript, "language_detected": lang_detected, }