dsh-voice-dictation

Push-to-talk voice dictation for DeepSeek Harness (DSH): a microphone button in the message composer records your voice, transcribes it through a local speech-to-text server, and drops the text straight into the message draft — hands-free prompting.

How it works

• Browser half — mic button in the composer; hold/click to record, audio goes to the host half on release.
• Host half — POST /voice/transcribe forwards audio to a local Parakeet STT HTTP server (default http://127.0.0.1:5093), with a local whisper CLI as fallback when the service is unreachable.
• Transcribed text lands in the composer draft; you review and send.

Everything stays on your machine — no cloud speech APIs.

Mic button in the composer	Dictating…	Text lands in the draft

Install

cd ~/.dsh/profiles/web
rm -rf node_modules/@deepseek-ai/dsh-voice-dictation   # drop any stale copy
pnpm add --force "file:/path/to/this/repo"

Add the composition row to ~/.dsh/profiles/web/cordis.patch.yml:

- insert:
    - id: voice-dictation
      name: '@deepseek-ai/dsh-voice-dictation'

Restart DSH web and look for the mic button in the composer.

Configuration

Key	Default	Meaning
sttUrl	http://127.0.0.1:5093	Parakeet-compatible STT HTTP endpoint
sttModel	server default	model selector passed through to the STT server

You'll need an STT backend running, e.g. NVIDIA Parakeet served over HTTP on the configured port. Without one, the plugin degrades to the whisper CLI fallback (requires whisper on PATH).

License

MIT — see LICENSE.