openspeechapi 0.2.9__tar.gz → 0.2.11__tar.gz

{openspeechapi-0.2.9 → openspeechapi-0.2.11}/.gitignore

@@ -12,6 +12,7 @@ dist/
 build/
 *.pyc
 *.pyo
+.DS_Store
 
 # macOS STT compiled artifacts
 scripts/engines/macos-stt/macos-stt-helper

openspeechapi-0.2.9/README.md → openspeechapi-0.2.11/PKG-INFO

@@ -1,3 +1,172 @@
+Metadata-Version: 2.4
+Name: openspeechapi
+Version: 0.2.11
+Summary: Unified speech interface for STT/TTS providers
+Requires-Python: >=3.11
+Requires-Dist: httpx>=0.27
+Requires-Dist: loguru>=0.7
+Requires-Dist: msgpack>=1.0
+Requires-Dist: pydantic>=2.0
+Requires-Dist: pyyaml>=6.0
+Provides-Extra: alibaba
+Provides-Extra: alibaba-stt
+Provides-Extra: alibaba-tts
+Provides-Extra: all
+Requires-Dist: elevenlabs; extra == 'all'
+Requires-Dist: faster-whisper; extra == 'all'
+Requires-Dist: openai; extra == 'all'
+Requires-Dist: openai-whisper; extra == 'all'
+Requires-Dist: piper-tts; extra == 'all'
+Requires-Dist: pyttsx3; (sys_platform == 'win32') and extra == 'all'
+Requires-Dist: torchaudio; extra == 'all'
+Requires-Dist: tts; extra == 'all'
+Requires-Dist: websockets; extra == 'all'
+Provides-Extra: assemblyai-stt
+Provides-Extra: audio
+Requires-Dist: numpy; extra == 'audio'
+Requires-Dist: sounddevice; extra == 'audio'
+Provides-Extra: azure
+Provides-Extra: azure-stt
+Provides-Extra: azure-tts
+Provides-Extra: baidu
+Provides-Extra: baidu-stt
+Provides-Extra: baidu-tts
+Provides-Extra: canary-qwen-stt
+Provides-Extra: cloud
+Requires-Dist: websockets; extra == 'cloud'
+Provides-Extra: coqui-tts
+Requires-Dist: tts; extra == 'coqui-tts'
+Provides-Extra: cosyvoice-tts
+Requires-Dist: torchaudio; extra == 'cosyvoice-tts'
+Provides-Extra: deepgram
+Requires-Dist: websockets; extra == 'deepgram'
+Provides-Extra: deepgram-stt
+Requires-Dist: websockets; extra == 'deepgram-stt'
+Provides-Extra: deepgram-tts
+Provides-Extra: dev
+Requires-Dist: numpy; extra == 'dev'
+Requires-Dist: pytest-asyncio>=0.24; extra == 'dev'
+Requires-Dist: pytest-cov; extra == 'dev'
+Requires-Dist: pytest-dotenv; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: ruff==0.15.*; extra == 'dev'
+Provides-Extra: dolphin-stt
+Requires-Dist: dataoceanai-dolphin; extra == 'dolphin-stt'
+Requires-Dist: torchcodec; extra == 'dolphin-stt'
+Provides-Extra: elevenlabs
+Requires-Dist: elevenlabs; extra == 'elevenlabs'
+Requires-Dist: websockets; extra == 'elevenlabs'
+Provides-Extra: elevenlabs-stt
+Requires-Dist: websockets; extra == 'elevenlabs-stt'
+Provides-Extra: elevenlabs-tts
+Requires-Dist: elevenlabs; extra == 'elevenlabs-tts'
+Provides-Extra: faster-whisper-stt
+Requires-Dist: faster-whisper; extra == 'faster-whisper-stt'
+Provides-Extra: fireredasr-stt
+Requires-Dist: fireredasr; extra == 'fireredasr-stt'
+Provides-Extra: fish-speech-tts
+Provides-Extra: funasr-stt
+Requires-Dist: funasr>=1.1.0; extra == 'funasr-stt'
+Requires-Dist: torch; extra == 'funasr-stt'
+Requires-Dist: torchaudio; extra == 'funasr-stt'
+Provides-Extra: gemma4-stt
+Requires-Dist: accelerate; (sys_platform != 'darwin') and extra == 'gemma4-stt'
+Requires-Dist: librosa; (sys_platform != 'darwin') and extra == 'gemma4-stt'
+Requires-Dist: mlx-vlm<0.6.2,>=0.6.1; (sys_platform == 'darwin') and extra == 'gemma4-stt'
+Requires-Dist: torch; (sys_platform != 'darwin') and extra == 'gemma4-stt'
+Requires-Dist: transformers; (sys_platform != 'darwin') and extra == 'gemma4-stt'
+Provides-Extra: google
+Provides-Extra: google-stt
+Provides-Extra: google-tts
+Provides-Extra: iflytek
+Requires-Dist: websockets; extra == 'iflytek'
+Provides-Extra: iflytek-stt
+Requires-Dist: websockets; extra == 'iflytek-stt'
+Provides-Extra: iflytek-tts
+Requires-Dist: websockets; extra == 'iflytek-tts'
+Provides-Extra: kimi-audio-stt
+Requires-Dist: torch; extra == 'kimi-audio-stt'
+Provides-Extra: macos-native
+Provides-Extra: minimax-tts
+Provides-Extra: mlx-whisper-stt
+Requires-Dist: mlx-whisper; (sys_platform == 'darwin') and extra == 'mlx-whisper-stt'
+Provides-Extra: mms-stt
+Requires-Dist: soundfile; extra == 'mms-stt'
+Requires-Dist: torch; extra == 'mms-stt'
+Requires-Dist: transformers; extra == 'mms-stt'
+Provides-Extra: moonshine-stt
+Requires-Dist: soundfile; extra == 'moonshine-stt'
+Requires-Dist: torch; extra == 'moonshine-stt'
+Requires-Dist: transformers; extra == 'moonshine-stt'
+Provides-Extra: openai
+Requires-Dist: openai; extra == 'openai'
+Provides-Extra: openai-stt
+Requires-Dist: openai; extra == 'openai-stt'
+Provides-Extra: openai-tts
+Requires-Dist: openai; extra == 'openai-tts'
+Provides-Extra: paraformer-stt
+Requires-Dist: funasr>=1.1.0; extra == 'paraformer-stt'
+Requires-Dist: torch; extra == 'paraformer-stt'
+Requires-Dist: torchaudio; extra == 'paraformer-stt'
+Provides-Extra: parakeet-stt
+Requires-Dist: parakeet-mlx; (sys_platform == 'darwin') and extra == 'parakeet-stt'
+Provides-Extra: phi4-multimodal-stt
+Requires-Dist: accelerate; extra == 'phi4-multimodal-stt'
+Requires-Dist: backoff; extra == 'phi4-multimodal-stt'
+Requires-Dist: peft; extra == 'phi4-multimodal-stt'
+Requires-Dist: pillow; extra == 'phi4-multimodal-stt'
+Requires-Dist: scipy; extra == 'phi4-multimodal-stt'
+Requires-Dist: soundfile; extra == 'phi4-multimodal-stt'
+Requires-Dist: torch; extra == 'phi4-multimodal-stt'
+Requires-Dist: torchvision; extra == 'phi4-multimodal-stt'
+Requires-Dist: transformers; extra == 'phi4-multimodal-stt'
+Provides-Extra: piper-tts
+Requires-Dist: piper-tts; extra == 'piper-tts'
+Provides-Extra: qwen3-asr-stt
+Requires-Dist: modelscope; extra == 'qwen3-asr-stt'
+Requires-Dist: qwen-asr; extra == 'qwen3-asr-stt'
+Provides-Extra: qwen3-omni-stt
+Requires-Dist: accelerate; extra == 'qwen3-omni-stt'
+Requires-Dist: qwen-omni-utils; extra == 'qwen3-omni-stt'
+Requires-Dist: torch; extra == 'qwen3-omni-stt'
+Requires-Dist: transformers; extra == 'qwen3-omni-stt'
+Provides-Extra: sensevoice-stt
+Requires-Dist: funasr>=1.1.0; extra == 'sensevoice-stt'
+Requires-Dist: torch; extra == 'sensevoice-stt'
+Requires-Dist: torchaudio; extra == 'sensevoice-stt'
+Provides-Extra: server
+Requires-Dist: fastapi; extra == 'server'
+Requires-Dist: python-multipart; extra == 'server'
+Requires-Dist: uvicorn; extra == 'server'
+Requires-Dist: websockets; extra == 'server'
+Provides-Extra: sherpa-onnx-stt
+Requires-Dist: websockets; extra == 'sherpa-onnx-stt'
+Provides-Extra: tencent
+Provides-Extra: tencent-stt
+Provides-Extra: tencent-tts
+Provides-Extra: tracing
+Requires-Dist: opentelemetry-api; extra == 'tracing'
+Requires-Dist: opentelemetry-sdk; extra == 'tracing'
+Provides-Extra: volcengine
+Provides-Extra: volcengine-stt
+Provides-Extra: volcengine-tts
+Provides-Extra: vosk-stt
+Requires-Dist: huggingface-hub; extra == 'vosk-stt'
+Requires-Dist: vosk; extra == 'vosk-stt'
+Provides-Extra: voxtral-stt
+Requires-Dist: accelerate; extra == 'voxtral-stt'
+Requires-Dist: mistral-common[audio]>=1.8.1; extra == 'voxtral-stt'
+Requires-Dist: torch; extra == 'voxtral-stt'
+Requires-Dist: transformers>=4.54.0; extra == 'voxtral-stt'
+Provides-Extra: wenet-stt
+Provides-Extra: whisper-stt
+Requires-Dist: openai-whisper; extra == 'whisper-stt'
+Provides-Extra: whisperlivekit-stt
+Requires-Dist: websockets; extra == 'whisperlivekit-stt'
+Provides-Extra: windows-native
+Requires-Dist: pyttsx3; (sys_platform == 'win32') and extra == 'windows-native'
+Description-Content-Type: text/markdown
+
 # OpenSpeechAPI
 
 > Unified speech interface for STT/TTS providers — one API, multiple backends.
@@ -8,22 +177,24 @@ OpenSpeechAPI 提供统一的语音接口，通过字符串指定 provider 即
 
 ### 安装
 
+**方式一 · 通过 PyPI 安装(直接使用)**
 ```bash
-# 安装全部 provider
-pip install -e ".[all]"
-
-# 或按需安装
-pip install -e ".[openai]"           # OpenAI Whisper STT + TTS
-pip install -e ".[faster-whisper]"   # 本地 faster-whisper STT
-pip install -e ".[openai,faster-whisper]"  # 指定多个
-
-# 仅核心包（不含任何 provider）
-pip install -e .
+pip install "openspeechapi[server]"           # 起 HTTP 服务 / WebUI 必须带 [server](fastapi/uvicorn)
+pip install "openspeechapi[server,openai]"    # 服务 + 指定 provider
+pip install "openspeechapi[server,all]"       # 服务 + 全部 provider
+pip install openspeechapi                      # 仅核心库(库模式;不含服务,也起不了 server)
+```
 
-# 开发环境
-pip install -e ".[dev]"
+**方式二 · 源码安装(开发,可编辑)**
+```bash
+git clone https://github.com/wingsfly/OpenSpeechAPI.git
+cd OpenSpeechAPI
+uv venv && uv pip install -e ".[server,dev]"  # 或 pip install -e ".[server,dev]";按需换 .[all] 等
 ```
 
+> ⚠️ 纯 `pip install openspeechapi`(核心库)**不含 fastapi/uvicorn**,无法 `serve`;起服务请带 `[server]`。
+> 两种方式启动服务的差异见下方 [启动服务](#启动服务)。
+
 ### 30 秒上手 — TTS
 
 ```python
@@ -205,7 +376,24 @@ python -m openspeechapi.demo tts -t "Hello world" --play \
 | `whisperlivekit-stt` | STT | WhisperLiveKit 本地服务（Deepgram 兼容 WS，支持 MLX 后端） | local | `pip install -e ".[whisperlivekit]"` |
 | `elevenlabs-stt` | STT | ElevenLabs Scribe API（云端，支持实时流式 WS + 批量） | remote | `pip install -e ".[elevenlabs-stt]"` |
 | `deepgram` | STT | Deepgram API（云端，支持实时流式） | remote | `pip install -e ".[deepgram]"` |
-| `gemma4` | STT | Google Gemma 4 多模态 ASR（macOS/MLX 本地，E4B 默认/12B 可选，>30s 自动分段，支持转写/翻译/理解） | subprocess | `pip install -e ".[gemma4-stt]"` |
+| `gemma4` | STT | Google Gemma 4 多模态 ASR（macOS/MLX 本地，E2B/E4B，>30s 自动分段；任务：转写 / 翻译(任意目标语言) / 理解 / 问答 / 语种识别） | subprocess | `pip install -e ".[gemma4-stt]"` |
+| `sensevoice` | STT | SenseVoice-Small 本地多语种 ASR（FunASR，zh/粤/en/ja/ko，比 Whisper 快 ~15-50×） | subprocess | `pip install -e ".[sensevoice-stt]"` |
+| `qwen3-asr` | STT | Qwen3-ASR 本地多语种 ASR（2026 开源 SOTA，中/方言/英，0.6B/1.7B） | subprocess | `pip install -e ".[qwen3-asr-stt]"` |
+| `mlx-whisper` | STT | Whisper on Apple MLX（本地，large-v3 / turbo，中/英多语种，仅 Apple Silicon） | subprocess | `pip install -e ".[mlx-whisper-stt]"` |
+| `paraformer` | STT | Paraformer 本地 ASR（FunASR，普通话 SOTA 级，VAD+标点，zh/en） | subprocess | `pip install -e ".[paraformer-stt]"` |
+| `funasr` | STT | FunASR 总入口（任选模型库 + VAD/标点/说话人分离） | subprocess | `pip install -e ".[funasr-stt]"` |
+| `fireredasr` | STT | 小红书 FireRedASR（普通话 SOTA+方言+英文，歌词识别，AED/LLM） | subprocess | `pip install -e ".[fireredasr-stt]"` |
+| `dolphin` | STT | DataoceanAI Dolphin（40 东方语种 + 22 中文方言，small/base） | subprocess | `pip install -e ".[dolphin-stt]"` |
+| `wenet` | STT | WeNet U2++ Conformer（生产级，zh/en 预置；流式后续） | subprocess | WebUI Engines 安装,或 `pip install 'wenet @ git+https://github.com/wenet-e2e/wenet.git'` |
+| `canary-qwen` | STT | NVIDIA Canary-Qwen-2.5B（Open ASR 英文第1，SALM；仅英文，需 NeMo+GPU） | subprocess | WebUI Engines 安装,或 `pip install 'nemo_toolkit[asr] @ git+https://github.com/NVIDIA/NeMo.git'` |
+| `parakeet` | STT | NVIDIA Parakeet-TDT on MLX（最快，v2 英文/v3 欧语；中文弱，仅 Apple Silicon） | subprocess | `pip install -e ".[parakeet-stt]"` |
+| `qwen3-omni` | STT | Qwen3-Omni-30B 全模态 LLM（ASR+理解，zh/en+；需大 GPU ~60GB） | subprocess | `pip install -e ".[qwen3-omni-stt]"` |
+| `voxtral` | STT | Mistral Voxtral（Mini-3B/Small-24B，转写+理解，多语种；建议 GPU） | subprocess | `pip install -e ".[voxtral-stt]"` |
+| `phi4-multimodal` | STT | 微软 Phi-4-multimodal（多模态 LLM，ASR+理解，zh/en+；建议 GPU） | subprocess | `pip install -e ".[phi4-multimodal-stt]"` |
+| `kimi-audio` | STT | 月之暗面 Kimi-Audio-7B（音频基础模型，ASR+理解，zh/en；需 Linux+CUDA/flash-attn） | subprocess | WebUI Engines 安装,或 `pip install 'kimi-audio @ git+https://github.com/MoonshotAI/Kimi-Audio.git'` |
+| `moonshine` | STT | Useful Sensors Moonshine（边缘/实时英文 ASR，tiny/base，轻量） | subprocess | `pip install -e ".[moonshine-stt]"` |
+| `vosk` | STT | Vosk（Kaldi 离线，20+ 语言含 zh/en，轻量低资源） | subprocess | `pip install -e ".[vosk-stt]"` |
+| `mms` | STT | Meta MMS（Wav2Vec2-CTC，1000+ 语言含 zh/en，按 ISO 639-3 选语言） | subprocess | `pip install -e ".[mms-stt]"` |
 | `openai-tts` | TTS | OpenAI Speech API（云端，支持流式） | remote | `pip install -e ".[openai]"` |
 | `elevenlabs` | TTS | ElevenLabs 高质量语音（云端，支持 HTTP/WS 流式） | remote | `pip install -e ".[elevenlabs-tts]"` |
 | `minimax` | TTS | Minimax 语音合成（云端） | remote | `pip install -e ".[minimax]"` |
@@ -229,6 +417,8 @@ print(list_providers())
 #  'whisperlivekit-stt']
 ```
 
+> **音频输入格式**:STT 上传支持 WAV/PCM/MP3/FLAC/OGG/WebM 等。引擎不能直接处理的格式会由服务端自动转为 16k 单声道 WAV(压缩格式需 `ffmpeg`);缺 ffmpeg 且格式不被支持时返回 400,Web UI 会在上传/录音前拦截提示。详见 [docs/architecture/audio-format-negotiation.md](docs/architecture/audio-format-negotiation.md)。
+
 ## Provider 参数
 
 ### `openai-stt`
@@ -263,6 +453,204 @@ create_provider("faster-whisper",
 )
 ```
 
+### `gemma4`
+
+```python
+create_provider("gemma4",
+    model="mlx-community/gemma-4-E4B-it-8bit",  # E2B/E4B（8bit 翻译更准；勿用 12B）
+    task="transcribe",          # transcribe｜translate｜understand｜qa｜detect_language
+    target_language="English",  # task=translate 的目标语言（任意语言）
+    include_transcript=False,   # task=translate：同时输出源转写 + 译文
+)
+```
+
+macOS / Apple Silicon 本地多模态 ASR（mlx-vlm）。5 个任务及全部字段可在 Web UI 的 Lab「Advanced Options」按请求覆盖。详见 [docs/architecture/gemma4-stt-provider.md](docs/architecture/gemma4-stt-provider.md)。
+
+### `sensevoice`
+
+```python
+create_provider("sensevoice",
+    model="FunAudioLLM/SenseVoiceSmall",
+    language="auto",      # auto|zh|en|yue|ja|ko|nospeech
+    device="cpu",         # cpu|mps|cuda
+    use_itn=True,         # 标点/数字规整
+)
+```
+
+FunASR 本地多语种 ASR（zh/粤/en/ja/ko），非自回归、极快；全部字段可在 Lab「Advanced Options」按请求覆盖。详见 [docs/architecture/sensevoice-stt-provider.md](docs/architecture/sensevoice-stt-provider.md)。
+
+### `qwen3-asr`
+
+```python
+create_provider("qwen3-asr",
+    model="Qwen/Qwen3-ASR-0.6B",  # 或 Qwen/Qwen3-ASR-1.7B
+    language="auto",              # auto|Chinese|English|Cantonese|Japanese|Korean
+    device="cpu",                 # cpu|mps|cuda
+)
+```
+
+阿里 Qwen3-ASR（2026 开源 ASR SOTA，中/方言/英）本地推理（qwen-asr 包）。需另装 `torch`。详见 [docs/architecture/qwen3-asr-stt-provider.md](docs/architecture/qwen3-asr-stt-provider.md)。
+
+### `mlx-whisper`
+
+```python
+create_provider("mlx-whisper",
+    model="mlx-community/whisper-large-v3-turbo",  # 或 whisper-large-v3-mlx
+    language="auto",                               # auto|en|zh|yue|ja|ko|...
+)
+```
+
+Apple Silicon 原生 Whisper（MLX），large-v3 / turbo，中英文多语种。仅 macOS/Apple Silicon。详见 [docs/architecture/mlx-whisper-stt-provider.md](docs/architecture/mlx-whisper-stt-provider.md)。
+
+### `paraformer`
+
+```python
+create_provider("paraformer",
+    model="funasr/paraformer-zh",  # 或 funasr/paraformer-en
+    vad=True, punc=True,           # VAD 切分 + 标点恢复
+)
+```
+
+阿里 Paraformer（FunASR），普通话 SOTA 级非自回归 ASR，带 VAD + 标点。详见 [docs/architecture/paraformer-stt-provider.md](docs/architecture/paraformer-stt-provider.md)。
+
+### `funasr`
+
+```python
+create_provider("funasr",
+    model="funasr/paraformer-zh",  # 模型库任意条目
+    spk=True,                      # CAM++ 说话人分离 → [spk0]/[spk1] 标注
+)
+```
+
+FunASR 通用总入口:任选模型库模型 + VAD/标点/**说话人分离**。详见 [docs/architecture/funasr-stt-provider.md](docs/architecture/funasr-stt-provider.md)。
+
+### `fireredasr`
+
+```python
+create_provider("fireredasr",
+    model_type="aed",   # aed(≤60s) | llm(≤30s);权重自动下载
+)
+```
+
+小红书 FireRedASR,普通话公开基准 SOTA + 方言 + 英文,歌词识别强。详见 [docs/architecture/fireredasr-stt-provider.md](docs/architecture/fireredasr-stt-provider.md)。
+
+### `dolphin`
+
+```python
+create_provider("dolphin",
+    model_name="small",   # small | base
+    lang_sym="zh", region_sym="CN",   # 留空则自动检测
+)
+```
+
+DataoceanAI Dolphin,40 种东方语言 + 22 种中文方言。详见 [docs/architecture/dolphin-stt-provider.md](docs/architecture/dolphin-stt-provider.md)。
+
+### `wenet`
+
+```python
+create_provider("wenet",
+    model="chinese",   # chinese | english
+)
+```
+
+WeNet 生产级 U2++ Conformer(zh/en 预置)。从 git 安装(不在 PyPI)。详见 [docs/architecture/wenet-stt-provider.md](docs/architecture/wenet-stt-provider.md)。
+
+### `canary-qwen`
+
+```python
+create_provider("canary-qwen",
+    model="nvidia/canary-qwen-2.5b",
+    device="cuda", dtype="bfloat16",   # 仅英文;强烈建议 GPU
+)
+```
+
+NVIDIA Canary-Qwen-2.5B(Open ASR 英文第 1,SALM)。**仅英文**;NeMo 重型安装 + 建议 GPU。详见 [docs/architecture/canary-qwen-stt-provider.md](docs/architecture/canary-qwen-stt-provider.md)。
+
+### `parakeet`
+
+```python
+create_provider("parakeet",
+    model="mlx-community/parakeet-tdt-0.6b-v2",  # v2 英文;v3 + 欧洲语言
+)
+```
+
+NVIDIA Parakeet-TDT on Apple MLX,榜上最快。英文/欧语为主,**中文弱**;仅 Apple Silicon。详见 [docs/architecture/parakeet-stt-provider.md](docs/architecture/parakeet-stt-provider.md)。
+
+### `qwen3-omni`
+
+```python
+create_provider("qwen3-omni",
+    model="Qwen/Qwen3-Omni-30B-A3B-Instruct",
+    prompt="Transcribe the audio into text.",   # 改成问题即可做音频问答
+)
+```
+
+阿里 Qwen3-Omni-30B-A3B 全模态 LLM(ASR + 音频理解,zh/en+)。**需大显存 GPU(~60GB),笔记本装不下**。详见 [docs/architecture/qwen3-omni-stt-provider.md](docs/architecture/qwen3-omni-stt-provider.md)。
+
+### `voxtral`
+
+```python
+create_provider("voxtral",
+    model="mistralai/Voxtral-Mini-3B-2507",  # 或 Voxtral-Small-24B-2507
+    language="en",
+)
+```
+
+Mistral Voxtral(转写 + 音频理解,多语种)。3B/24B,建议 GPU。详见 [docs/architecture/voxtral-stt-provider.md](docs/architecture/voxtral-stt-provider.md)。
+
+### `phi4-multimodal`
+
+```python
+create_provider("phi4-multimodal",
+    model="microsoft/Phi-4-multimodal-instruct",
+    prompt="Transcribe the audio clip into text.",
+)
+```
+
+微软 Phi-4-multimodal,紧凑多模态 LLM(ASR + 音频理解,zh/en+)。建议 GPU。详见 [docs/architecture/phi4-multimodal-stt-provider.md](docs/architecture/phi4-multimodal-stt-provider.md)。
+
+### `kimi-audio`
+
+```python
+create_provider("kimi-audio",
+    model="moonshotai/Kimi-Audio-7B-Instruct",
+    prompt="Please transcribe the audio into text.",
+)
+```
+
+月之暗面 Kimi-Audio-7B 音频基础模型(ASR + 音频理解,zh/en)。git 安装 + 建议 GPU。详见 [docs/architecture/kimi-audio-stt-provider.md](docs/architecture/kimi-audio-stt-provider.md)。
+
+### `moonshine`
+
+```python
+create_provider("moonshine",
+    model="UsefulSensors/moonshine-base",  # base | tiny
+)
+```
+
+Useful Sensors Moonshine,边缘/实时英文 ASR,轻量快速。详见 [docs/architecture/moonshine-stt-provider.md](docs/architecture/moonshine-stt-provider.md)。
+
+### `vosk`
+
+```python
+create_provider("vosk",
+    model="vosk-model-small-en-us-0.15",  # 中文: vosk-model-small-cn-0.22
+)
+```
+
+Vosk(Kaldi 离线),20+ 语言,轻量低资源,模型自动下载。详见 [docs/architecture/vosk-stt-provider.md](docs/architecture/vosk-stt-provider.md)。
+
+### `mms`
+
+```python
+create_provider("mms",
+    model="facebook/mms-1b-all",
+    language="eng",   # ISO 639-3: eng / cmn / yue / jpn ...
+)
+```
+
+Meta MMS(Wav2Vec2-CTC),1000+ 语言含中英,按 **ISO 639-3** 码切换语言适配器;CTC 输出小写无标点。详见 [docs/architecture/mms-stt-provider.md](docs/architecture/mms-stt-provider.md)。
+
 ### `openai-tts`
 
 ```python
@@ -407,10 +795,25 @@ bash scripts/engines/macos-stt/install.sh
 
 ### 启动服务
 
+**pip 安装后**(已带 `[server]`)—— 配置自动解析/生成,开箱即起:
 ```bash
-openspeechapi serve --config providers.yaml --port 8600
+openspeechapi serve                 # 自动解析配置;没有则生成默认(macOS 默认 macos_tts)
+openspeechapi serve --port 8600     # 指定端口
 ```
 
+**源码目录运行**:
+```bash
+python -m openspeechapi.cli serve   # 或 openspeechapi serve;在仓库目录优先用 ./providers.yaml
+```
+
+启动后打开 WebUI:**http://127.0.0.1:8600/ui/**
+
+**配置解析顺序**(`--config` 可放在 `serve` **前或后**,例如 `openspeechapi --config x serve` 或 `openspeechapi serve --config x`):
+1. 显式 `--config <path>`
+2. 当前目录 `./providers.yaml`(源码目录运行时优先)
+3. `~/.config/openspeechapi/providers.yaml`(遵循 `XDG_CONFIG_HOME`)
+4. 都没有 → 在 `~/.config/openspeechapi/providers.yaml` **自动生成**一份可用默认配置
+
 ### Python Client（与 Library 模式接口一致）
 
 ```python