Feature: Add SenseVoice/FunASR as an alternative STT backend

Proposal

Consider SenseVoiceSmall / FunASR as an optional self-hosted STT backend for VibeVoice workflows that need a compact, non-autoregressive recognizer.

SenseVoiceSmall is a 234M-parameter model with:

• speech recognition for Mandarin, Cantonese, English, Japanese, and Korean;
• language, speech-emotion, and audio-event tags;
• non-autoregressive decoding;
• local CPU or GPU inference through FunASR.

For long recordings, FunASR can compose SenseVoiceSmall with the separate FSMN-VAD model. VAD is a pipeline component, not a built-in part of the SenseVoice checkpoint.

Integration sketch

from funasr import AutoModel

model = AutoModel(
    model="iic/SenseVoiceSmall",
    vad_model="fsmn-vad",
    trust_remote_code=True,
)
result = model.generate(input=audio_path)

This could be exposed as an optional backend rather than replacing VibeVoice's own ASR path, keeping the integration narrow and useful for local CJK transcription.

References

• SenseVoice: https://github.com/FunAudioLLM/SenseVoice
• SenseVoiceSmall model card: https://huggingface.co/FunAudioLLM/SenseVoiceSmall
• FunASR: https://github.com/modelscope/FunASR
• Paper: https://arxiv.org/abs/2407.04051