Feature: Add SenseVoice as STT module — 5x faster, non-autoregressive

[LauraGPT]

Hi! Verbi's modular architecture is perfect for experimenting with different STT backends.

I'd like to suggest SenseVoice as a transcription module option:

Why SenseVoice?

• 5x faster than Whisper — non-autoregressive (single forward pass)
• 234M params — lightweight, can run alongside your LLM
• 50+ languages — auto-detection
• Emotion detection — useful for more natural voice assistant responses
• OpenAI-compatible API — minimal integration effort

Integration (fits your modular design)

from funasr import AutoModel

class SenseVoiceTranscriber:
    def __init__(self):
        self.model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
    
    def transcribe(self, audio):
        result = self.model.generate(input=audio)
        return result[0]["text"]

Links

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