Feature request: Add FunASR/SenseVoice as audio model backend

## Summary Xinference supports various LLM/embedding/image/audio models. Would you consider adding [FunASR](https://github.com/modelscope/FunASR) models (SenseVoice, Paraformer, Fun-ASR-Nano) as audio/speech model backends? ## Why FunASR? FunASR is the most popular open-source ASR toolkit for Chinese and multilingual speech recognition: - **SenseVoice** (234M params): Non-autoregressive, ~25x faster than Whisper-large, 50+ languages, emotion + audio event detection - **Paraformer** (220M params): ~170x realtime on GPU for Chinese, built-in VAD + punctuation - **Fun-ASR-Nano** (800M params): LLM-based ASR (SenseVoice encoder + Qwen3-0.6B decoder), 31 languages - **cam++**: Speaker diarization model (7.2M params) ## Integration FunASR already provides an OpenAI-compatible API server: ```bash pip install funasr vllm funasr-server --device cuda # http://localhost:8000/v1/audio/transcriptions ``` This could serve as a reference for integrating into Xinference's model serving framework. ## Model ecosystem | Model | Task | Params | Speed | |-------|------|--------|-------| | SenseVoice-Small | ASR + emotion | 234M | ~25x vs Whisper | | Paraformer-large | Chinese ASR | 220M | ~170x realtime | | Fun-ASR-Nano | Multilingual ASR | 800M | LLM-based | | FSMN-VAD | Voice Activity Detection | 0.4M | — | | CT-Punc | Punctuation | — | — | | cam++ | Speaker Diarization | 7.2M | — | All models available on ModelScope and HuggingFace (FunAudioLLM org). ## References - FunASR: https://github.com/modelscope/FunASR (16K+ stars) - SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K+ stars) - Install: `pip install funasr`