"""音频相关 API 端点 — TTS（文本转语音）、ASR（语音转文本）等。 音频端点使用神识的模型注册表查找可用的音频模型，并调用对应音频适配器合成 / 转写。ASR 与 TTS 按 ``tags`` 中的 ``asr`` / ``tts`` 能力区分，ASR 内部再按 ``provider`` 区分协议：``yuanqi`` 走 multipart ``/v1/audio/transcriptions``， 其他走 MiMo 风格的 ``/chat/completions`` + ``input_audio``。 """ from __future__ import annotations from fastapi import APIRouter, File, Form, HTTPException, Response, UploadFile from pydantic import BaseModel, Field from app.ai_provider.audio import AudioService, AudioServiceError from app.ai_provider.manager import ProviderManagerFactory from app.core.logging import logger router = APIRouter(tags=["audio"]) _MIME_MAP = { "wav": "audio/wav", "mp3": "audio/mpeg", "opus": "audio/opus", "flac": "audio/flac", } class _AudioModelFactory: """Adapter: wraps ProviderManager to provide ModelFactory interface for AudioService. ``create_audio(name)`` 按 ``tags`` 判断 ``asr`` / ``tts`` 能力： - ``tts`` → ``HunjingTTSOpenAI``（``/chat/completions`` + ``audio``） - ``asr`` → 按 ``provider`` 分派： - ``yuanqi`` → ``YuanqiASRWhisper``（multipart ``/v1/audio/transcriptions``） - 其他 → ``HunjingASROpenAI``（``/chat/completions`` + ``input_audio``） """ def __init__(self, manager) -> None: self._manager = manager @property def registry(self): return self._manager.model_registry def create_audio(self, name: str): model = self._manager.get_model_config(name) is_asr = "asr" in model.tags if is_asr: return self._create_asr(model) return self._create_tts(model) def _provider_or_none(self, model): """安全获取 provider 配置（provider 未配置时返回 None，走纯端点解析）。""" try: return self._manager.get_provider_config(model.provider) except Exception: return None # ── ASR ───────────────────────────────────────────────────────────── def _create_asr(self, model): """按 provider 选择 ASR 协议适配器。""" # yuanqi 本地引擎走 multipart /v1/audio/transcriptions if model.provider == "yuanqi": from app.ai_provider.audio.yuanqi_asr import YuanqiASRWhisper if model.endpoints: return YuanqiASRWhisper.from_model_config( model, provider=self._provider_or_none(model) ) provider = self._manager.get_provider_config(model.provider) api_key, base_url = self._resolve_credentials(provider) if not base_url: raise AudioServiceError( status_code=500, code="audio_asr_no_base_url", message=f"Provider '{model.provider}' has no base_url configured.", ) return YuanqiASRWhisper( base_url=base_url, api_key=api_key or "", model=model.model, request_timeout_seconds=model.timeout_seconds or 300, ) # MiMo 等走 /chat/completions + input_audio from app.ai_provider.audio.asr import HunjingASROpenAI if model.endpoints: return HunjingASROpenAI.from_model_config( model, provider=self._provider_or_none(model) ) provider = self._manager.get_provider_config(model.provider) api_key, base_url = self._resolve_credentials(provider) if not base_url: raise AudioServiceError( status_code=500, code="audio_asr_no_base_url", message=f"Provider '{model.provider}' has no base_url configured.", ) return HunjingASROpenAI( base_url=base_url, api_key=api_key or "", model=model.model, request_timeout_seconds=model.timeout_seconds or 300, ) # ── TTS ───────────────────────────────────────────────────────────── def _create_tts(self, model): from app.ai_provider.audio.tts import HunjingTTSOpenAI # 有 endpoints 时交给 from_model_config（按端点的 status/model/超时） if model.endpoints: return HunjingTTSOpenAI.from_model_config( model, provider=self._provider_or_none(model) ) # 无 endpoint 时回退到 provider 级别凭证 provider = self._manager.get_provider_config(model.provider) api_key, base_url = self._resolve_credentials(provider) if not base_url: raise AudioServiceError( status_code=500, code="audio_tts_no_base_url", message=f"Provider '{model.provider}' has no base_url configured.", ) return HunjingTTSOpenAI( base_url=base_url, api_key=api_key or "", model=model.model, request_timeout_seconds=model.timeout_seconds or 120, ) def _resolve_credentials(self, provider): """从 settings 解析 provider 级别的 API key 和 base URL。""" from app.core.config import settings api_key = ( settings.__getattribute__(provider.api_key_env.lower()) if hasattr(settings, provider.api_key_env.lower()) else "" ) base_url = ( settings.__getattribute__(provider.base_url_env.lower()) if hasattr(settings, provider.base_url_env.lower()) else None ) return api_key, base_url class TTSRequest(BaseModel): """TTS 合成请求。""" text: str = Field(..., min_length=1, max_length=4096, description="要合成语音的文本。") voice: str = Field("default", description="语音角色标识（取决于模型，如 default 或声纹 ID）。") format: str = Field("wav", description="音频格式：wav / mp3 / opus / flac。") @router.post("/audio/tts") async def synthesize_speech(body: TTSRequest) -> Response: """合成文本为语音，返回音频二进制流。 使用神识注册表中第一个活跃的 TTS 模型（capability: tts）。 未配置 TTS 模型时返回 503。 """ if not body.text.strip(): raise HTTPException(status_code=400, detail="text 不能为空。") mime = _MIME_MAP.get(body.format, "audio/wav") try: manager = ProviderManagerFactory.create() service = AudioService(_AudioModelFactory(manager)) async with service: result = await service.synthesize( text=body.text, voice=body.voice, response_format=body.format, ) except AudioServiceError as exc: logger.error( "audio.tts.api_error", voice=body.voice, format=body.format, error_code=exc.code, error_message=exc.message, ) raise HTTPException( status_code=exc.status_code or 502, detail=exc.message, ) from exc logger.info( "audio.tts.success", voice=body.voice, format=body.format, audio_bytes=len(result.audio_data), ) return Response( content=result.audio_data, media_type=mime, headers={ "Content-Disposition": f'inline; filename="speech.{body.format}"', "X-Audio-Format": body.format, "Cache-Control": "public, max-age=3600", }, ) # ── ASR（语音转文本）─────────────────────────────────────────────────────── class TranscriptionResponse(BaseModel): """OpenAI Whisper 兼容的转写响应。""" text: str = Field(description="转写后的文本") @router.post("/audio/transcriptions", response_model=TranscriptionResponse) async def transcribe_audio( file: UploadFile = File(..., description="音频文件（wav/mp3/m4a/flac/ogg/opus/webm 等）"), language: str | None = Form(default=None, description="语言提示（ISO-639-1：zh/en/ja/ko 等），省略则自动检测"), model: str | None = Form(default=None, description="可选，指定使用的 ASR 模型逻辑名（如 funasr-sensevoice / mimo-v2.5-asr），省略则用注册表中第一个"), ) -> TranscriptionResponse: """转写音频为文本，返回 OpenAI Whisper 兼容的 JSON。 不指定 ``model`` 时用神识注册表中第一个活跃的 ASR 模型（capability: asr）， 按其 ``provider`` 自动选择协议适配器：``yuanqi`` 走本地 multipart 端点， 其他走 MiMo 风格的 chat completions。未配置 ASR 模型时返回 503。 """ audio_bytes = await file.read() if not audio_bytes: raise HTTPException(status_code=400, detail="音频文件为空。") fmt = (file.filename or "audio.wav").rsplit(".", 1)[-1].lower() if "." in (file.filename or "") else "wav" try: manager = ProviderManagerFactory.create() service = AudioService(_AudioModelFactory(manager)) async with service: result = await service.transcribe( audio_data=audio_bytes, format=fmt, language=language, model_name=model, ) except AudioServiceError as exc: logger.error( "audio.asr.api_error", filename=file.filename, format=fmt, language=language, error_code=exc.code, error_message=exc.message, ) raise HTTPException( status_code=exc.status_code or 502, detail=exc.message, ) from exc logger.info( "audio.asr.success", filename=file.filename, format=fmt, language=language, text_length=len(result.text), text=result.text, ) return TranscriptionResponse(text=result.text)