Feature request: Add FunASR/Paraformer as alternative ASR backend

[LauraGPT]

Feature Request

Would it be possible to add FunASR's Paraformer model as an alternative ASR backend alongside Whisper?

Why

• Paraformer is a non-autoregressive model — it runs significantly faster than autoregressive Whisper, especially on long audio (170x realtime on GPU)
• For Chinese/Japanese/Korean audio, Paraformer and SenseVoice consistently outperform Whisper (see SenseVoice benchmark)
• Fun-ASR-Nano (LLM-based, 800M params) supports 31 languages with word-level timestamps
• FunASR models are available as ONNX via Sherpa-ONNX, which faster-whisper could potentially integrate

FunASR Models

Model	Params	Languages	Speed	Timestamps
SenseVoice-Small	234M	50+	25x realtime (CPU)	✅
Paraformer-large	220M	zh/en/ja/ko/yue	170x realtime (GPU)	✅
Fun-ASR-Nano	800M	31	GPU via vLLM	✅ word-level

Quick test

pip install funasr

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="audio.wav")

References

• FunASR GitHub (16K+ stars)
• SenseVoice (8K+ stars)
• Fun-ASR-Nano — LLM-based ASR