feat: Add FunASR/SenseVoice as speech recognition model backend

[LauraGPT]

Feature request

Xinference already supports various model types (LLM, embedding, image, audio). Adding FunASR/SenseVoice as a speech recognition backend would complement the existing audio capabilities.

Why FunASR for Xinference

• 5x faster than Whisper (SenseVoice is non-autoregressive, 234M params)
• Full pipeline: ASR + VAD + punctuation + speaker diarization in one toolkit
• 50+ languages with automatic language detection
• OpenAI-compatible API — funasr-server at /v1/audio/transcriptions
• Multiple models: SenseVoice (fast), Paraformer (Chinese-optimized), Fun-ASR-Nano (LLM-based, 31 languages)
• 1M+ monthly pip installs — widely adopted

Integration approach

FunASR models could be registered as a new model type in Xinference:

# Model registration
{
    "model_name": "SenseVoice-Small",
    "model_type": "audio",
    "model_family": "funasr",
    "model_id": "iic/SenseVoiceSmall",
}

The inference backend would wrap FunASR's Python API:

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
result = model.generate(input=audio_data)

References

• FunASR: https://github.com/modelscope/FunASR (16K+ stars)
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K+ stars)
• PyPI: pip install funasr