Feature: Add SenseVoice/FunASR as alternative STT backend — non-autoregressive, 5x faster

[LauraGPT]

Hi! VibeVoice is impressive — frontier voice AI with full pipeline.

I'd like to suggest SenseVoice from FunASR as an alternative STT option. It offers a fundamentally different architecture that avoids some issues seen with autoregressive ASR (like the "infinite loop" in #373):

Key differences

	Autoregressive ASR	SenseVoice (Non-autoregressive)
Decoding	Sequential token generation	Single forward pass
Failure mode	Can loop/degenerate	Deterministic output
Speed	Grows with input length	Constant time per segment
Params	Varies	234M
RTF (GPU)	~0.1	~0.006 (170x realtime)

The non-autoregressive architecture fundamentally cannot produce "infinite loop" outputs since it decodes the entire sequence in one step.

Features

• 50+ languages with auto-detection
• Emotion detection — happy/sad/angry/neutral
• Audio event classification — laughter, music, applause
• Built-in VAD (FSMN-VAD, 5.2M params) + punctuation restoration

Integration

from funasr import AutoModel

model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
result = model.generate(input=audio_chunk)
# Returns: text + language + emotion + events

OpenAI-compatible server:

pip install funasr && funasr-server --device cuda
# POST /v1/audio/transcriptions

Links

• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)
• FunASR: https://github.com/modelscope/FunASR (16.7K stars)
• Paper: https://arxiv.org/abs/2407.04051

This could serve as a lightweight, fast alternative for users who don't need the full VibeVoice-ASR model but want reliable, fast transcription.