Feature Request: Add FunASR as an ASR backend option

[LauraGPT]

Summary

Would it be possible to add FunASR as an alternative ASR backend alongside faster-whisper?

Why FunASR?

FunASR's toolkit source is MIT-licensed; individual model weights retain their own model-card licenses. It has some compelling advantages:

	FunASR (SenseVoice)	faster-whisper large-v3
Speed (GPU)	170x realtime	~46x realtime
Speed (CPU)	17x realtime	not viable
Speaker diarization	Built-in (cam++)	Needs pyannote
Punctuation	Built-in	Needs post-processing
Languages	50+	57
Streaming	WebSocket native	Needs wrapper
License	Toolkit code: MIT; weights vary (SenseVoiceSmall: custom license)	MIT

Integration approach

FunASR already provides an OpenAI-compatible API endpoint (funasr-server), so it could potentially be integrated as:

1. Backend option: similar to how faster-whisper is used — pip install funasr, then use AutoModel for inference
2. Proxy mode: point speaches at a running funasr-server instance

Basic usage:

from funasr import AutoModel

model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad", device="cuda")
result = model.generate(input="audio.wav")
# Returns: [{"text": "...", "timestamp": [...]}]

What I can help with

I'm a contributor to FunASR and happy to help with a PR if there's interest. The main effort would be writing a backend adapter that maps FunASR's output format to speaches' internal representation.

PyPI: pip install funasr (latest: 1.3.8)

[LauraGPT]

Opened #658, which adds FunASR (SenseVoice / Paraformer) as an optional, OpenAI-compatible transcription backend — lazy-imported so it doesn't affect the base install. Tested end-to-end with SenseVoiceSmall. Would appreciate a review whenever you have a chance! 🙏

[BeZazz]

Spam the place up much?