Overview
dsh-hold-to-talk
Hold-to-talk voice input for the DeepSeek Harness Web composer: hold the mouse on the input box, speak, see the running transcript float above it, release to insert the text into the draft. Recognition runs locally (SenseVoice via sherpa-onnx) — no API key, offline, audio never leaves the machine. 微信电脑版同款长按语音输入:在输入框上按住鼠标说话,浮层里边说边出字,松手把文字写进输入框,上滑取消;识别在本机跑,免密钥、离线、音频不出本机。
README / EN
Package documentation
Registry summary
Hold-to-talk voice input for the DeepSeek Harness Web composer: hold the mouse on the input box, speak, see the running transcript float above it, release to insert the text into the draft. Recognition runs locally (SenseVoice via sherpa-onnx) — no API key, offline, audio never leaves the machine. 微信电脑版同款长按语音输入:在输入框上按住鼠标说话,浮层里边说边出字,松手把文字写进输入框,上滑取消;识别在本机跑,免密钥、离线、音频不出本机。
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
- SenseVoice is **not** a streaming model. "Live" is a re-decode of the tail window every 1.5 s, so the preview trails speech by roughly 1.5–2 s; the final result (on release) always uses the complete audio. - Long dictation takes a moment to finish: ~3 s for 20 s of speech, ~9 s at the 60 s cap. - **Desktop mouse only.** Touch long-press is deliberately not implemented — it collides with text selection and the context menu on mobile. - The model download is 228 MB on first use.