dsh-voice-scribe

Polishing long prompts in DeepSeek Harness rarely calls for a keyboard as the only option. dsh-voice-scribe makes voice input as natural as typing: place the cursor in the composer, hold Alt (or tap the mic icon) and start speaking, then release to insert the transcribed text right where the cursor sits. No app switching, no copy-paste.

The recognition engine defaults to local offline SenseVoice, keeping audio on the device with zero config and zero API keys; the model (about 230 MB) auto-downloads once via a domestic mirror. If local is unavailable it falls back to browser Web Speech, and you can optionally configure multiple OpenAI-compatible endpoints in a server-side ASR chain that fails over automatically. Five languages are supported: Chinese, English, Cantonese, Japanese, and Korean. A hot-word table fixes misheared names and jargon. Before any LLM polish, a local rule pass strips filler words and collapses extra spaces, saving tokens; if the LLM call fails the raw transcript is preserved. Live intermediate results and a level meter appear while recording, and switching windows auto-cancels to prevent accidental captures.

It is built for users who treat voice as their primary input and care about privacy and zero-friction setup: no API keys to manage, no audio leaving the machine, no config to write. It complements dsh-better-input rather than replacing it—pick this one for a focused, private voice-input experience; pick the other if you also want prompt optimization and file-to-Markdown conversion. The two can coexist.