MyVoiceTyping: local-first Chinese voice typing app using FunASR/SenseVoice on macOS

[botaruibo]

Hi FunASR community, I’m one of the maintainers of MyVoiceTyping, an open-source macOS app that turns speech into text directly in the current input field.

Repo: https://github.com/botaruibo/MyVoiceTyping
Text polishing model: https://modelscope.cn/models/botaruibo/MyVoiceTyping-1.5b-q4
Fine-tuning dataset: https://github.com/botaruibo/MyVoiceTyping-Dataset

The project uses the FunASR / SenseVoice ecosystem as the local ASR foundation, together with local punctuation restoration and a small local GGUF text-polishing model. The goal is not just file transcription, but day-to-day Chinese voice typing: press a hotkey, speak, and paste cleaned-up text into chat, docs, notes, or issue trackers.

Why we built it this way:

• Privacy / data safety: audio and text can stay on the Mac by default, which is important for personal notes, workplace content, and sensitive drafts.
• Chinese-first workflow: optimized around Mandarin voice input, punctuation restoration, hot words, and lightweight cleanup after recognition.
• Open and replaceable: the app, model choice, and tuning dataset are public, so users can inspect or swap the local pipeline.
• Practical packaging: we are working on making local ASR + punctuation + polishing usable as a normal macOS input utility, not only as research demos or scripts.

It can be compared with tools like Typless, but the positioning is different: MyVoiceTyping is an early open-source, local-first alternative rather than a cloud product.

Would love feedback from FunASR users on latency, local deployment, packaging, and Chinese dictation quality.

[LauraGPT]

这个方向很实用。我核对了 release-0.02，链路确实是 SenseVoiceSmall ONNX、CT-Transformer 标点和本地 GGUF 修整。两个小建议：补一个明确的 LICENSE（GitHub 目前没有识别到许可证），再在 README 加一张 Apple 芯片、端到端延迟和峰值内存表；这样我们后续做 FunASR 生态案例推荐时可以更放心地引用。

[botaruibo]

Small follow-up: I published a Chinese article explaining why I’m positioning MyVoiceTyping less as plain dictation and more as a prompt buffer for AI coding workflows:

https://juejin.cn/post/7663165474498740262

The key idea is: speak → local ASR / punctuation / light polishing → reviewable text → then send to Codex / Claude Code / Cursor, instead of auto-submitting raw dictation. This is especially relevant for Chinese / mixed Chinese-English prompts, where project names, API paths, and technical terms need review before they are sent to an agent.

This is also why the project keeps emphasizing local-first processing, public app/model/dataset assets, and user-confirmed local history for future Self-Evolution.