Feature: Add SenseVoice for local dictation

[LauraGPT]

Note

License and capability clarification (2026-07-14): FunASR is a toolkit, not a single checkpoint. The FunASR and SenseVoice repository source code is MIT; model weights follow each model card. SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean, and its weights use the linked FunASR Model Open Source License Agreement. Fun-ASR-Nano-2512 is Apache-2.0. Language coverage, punctuation, and performance depend on the selected model and runtime configuration.

Hi! VoiceFlow is great for local voice dictation on Windows + Linux.

SenseVoice could improve dictation speed significantly:

Why SenseVoice?

• 5x faster than Whisper — non-autoregressive
• 234M params — lighter, faster startup, less RAM
• Built-in punctuation — no separate step needed
• 50+ languages — auto-detection
• Hold-to-talk friendly — constant-time decoding means instant results when you release

Integration

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad", punc_model="ct-punc")
result = model.generate(input=audio)
text = result[0]["text"]  # Ready-to-use punctuated text

Links

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

[infiniV]

Thanks for the suggestion — SenseVoice looks genuinely interesting, especially the non-autoregressive speed and built-in punctuation for a hold-to-talk flow. The catch is VoiceFlow is built on faster-whisper / CTranslate2, so FunASR means a second runtime and model-management path to maintain. Not a no — I'll evaluate it as an optional backend. Keeping this open to track.

[botaruibo]

+1 on SenseVoice for local dictation, especially for Chinese / Chinese-English mixed input.

I maintain a small macOS local-first voice typing project that uses a similar direction: local ASR, punctuation restoration, lightweight correction / polishing, then paste into the active input field: https://github.com/botaruibo/MyVoiceTyping

A few practical notes from that side:

• hold-to-talk works well for desktop dictation because users expect the final text right after release;
• SenseVoice-style fast local ASR is a better fit than cloud Web Speech APIs when privacy is part of the product promise;
• for Chinese input, punctuation and light post-ASR correction matter almost as much as raw recognition accuracy;
• in real typing workflows, the target is not just a transcript, but text that can be pasted directly into chat, docs, issues, or coding prompts.

The app is still early, but the related polishing model and tuning dataset are public, in case they are useful as references for local dictation UX / post-ASR cleanup:

• Model: https://modelscope.cn/models/botaruibo/MyVoiceTyping-1.5b-q4
• Dataset: https://github.com/botaruibo/MyVoiceTyping-Dataset

[infiniV]

Thanks @botaruibo, useful notes, and nice project. Agreed on all four points, especially that the target is paste-ready text rather than a raw transcript. Punctuation and light cleanup are most of what makes hold-to-talk actually feel good.

The tradeoff on my side is unchanged (FunASR means a second runtime next to CTranslate2), but a pluggable backend is the direction this is heading anyway, and proper CJK / mixed-language input is a solid reason to move it up. Bookmarking your polishing model and dataset too, post-ASR cleanup is something I want regardless of which engine is underneath.

Keeping this open to track.