Feature request: Add SenseVoice/Paraformer as alternative ASR backend

[LauraGPT]

Summary

Would it be possible to add SenseVoice and Paraformer as alternative ASR backends alongside the current CTranslate2-based Whisper?

Motivation

faster-whisper excels at efficient Whisper inference. However, some use cases would benefit from alternative ASR architectures:

• SenseVoice (234M params): Non-autoregressive model achieving ~25x faster than Whisper-large with comparable accuracy on 50+ languages. Also provides emotion detection and audio event classification.
• Paraformer: Non-autoregressive Chinese ASR with state-of-the-art accuracy on AISHELL benchmarks, including built-in VAD and punctuation.
• Fun-ASR-Nano: LLM-based ASR (SenseVoice encoder + Qwen3-0.6B decoder, 800M params) supporting 31 languages.

All models are available via FunASR (pip install funasr) and as ONNX exports via Sherpa-ONNX for CTranslate2-like optimized inference.

Benchmark comparison

Model	Type	Speed (GPU)	Languages	Extra features
Whisper-large-v3	Autoregressive	1x baseline	99	Translation
SenseVoice-Small	Non-AR	~25x	50+	Emotion, events
Paraformer-large	Non-AR	~170x realtime	Chinese	VAD, punctuation

Quick start

pip install funasr

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

References

• FunASR: https://github.com/modelscope/FunASR (16K+ stars)
• SenseVoice paper: https://arxiv.org/abs/2407.04051
• Paraformer paper: https://arxiv.org/abs/2206.08317