Feature: Add SenseVoice/FunASR as a local STT engine

[LauraGPT]

Important

Correction: The earlier comparative claims below are withdrawn; they were not established by a matched benchmark. SenseVoiceSmall supports Mandarin Chinese, Cantonese, English, Japanese, and Korean. It can emit language, emotion, and audio-event tags, but speaker diarization requires a separate model or pipeline (for example CAM++) and is not a built-in SenseVoice result. Runtime, timestamps, punctuation, and performance depend on the selected model, interface, hardware, and audio. FunASR and SenseVoice repository source code is MIT; model weights follow each model card. Please evaluate the exact integration on this project's workload.

Hi! AI Runner is great — offline inference with real-time voice conversations.

For the STT component, SenseVoice could improve voice conversation latency:

Why SenseVoice?

• 5x faster than Whisper — non-autoregressive
• 100% offline — aligns with your offline-first approach
• 234M params — lightweight, leaves resources for LLM + art generation
• Emotion detection — adapt conversation style based on user's tone
• Built-in VAD — handles speech boundaries automatically

Integration

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad")
result = model.generate(input=audio_chunk)
text = result[0]["text"]

Links

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