Feature: Add SenseVoice as ASR backend — 5x faster than faster-whisper

[LauraGPT]

Hi! Cool project — real-time voice assistant with WebRTC streaming.

I noticed you're using faster-whisper for ASR. Have you considered SenseVoice as an alternative backend? It could meaningfully reduce latency:

Performance comparison

Metric	faster-whisper (large-v3)	SenseVoice
Architecture	Autoregressive	Non-autoregressive
Relative speed	~4x vs original Whisper	~20x vs original Whisper
Model size	1.5B	234M
VRAM usage	~4-6GB	~1GB
First-token latency	Higher (sequential decoding)	Lower (parallel decoding)

For a real-time voice assistant, the non-autoregressive architecture gives much lower first-token latency — you get the full transcription in one forward pass instead of waiting for sequential token generation.

Integration options

Python API (drop-in):

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

OpenAI-compatible server (zero code change if you already use OpenAI STT API):

pip install funasr
funasr-server --device cuda
# POST /v1/audio/transcriptions — same as OpenAI/Whisper API

Links

• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)
• FunASR: https://github.com/modelscope/FunASR (16.7K stars)

Happy to discuss integration details!

[mogwai]

Interesting, should definitely be able to add this as a backend. Currently fswhisper is useful because it produces timings for words which then can be used to prefill the model