dsh-voice-input · Voice Input Plugin
中文
A voice-input plugin for DSH (DeepSeek Harness): adds a 🎤 mic button to the chat composer.
It uses the browser-native Web Speech API (zh-CN) to transcribe Chinese speech in real time;
when you finish speaking, the transcript is filled into the input draft for review/editing
before you press Enter to send. Zero config, zero keys, pure browser capability (Chrome / Edge).
Features
- Mic button on the right side of the composer (
conversation.input.right slot) — click to start listening, click again or pause to stop;
- Live transcript with a pulsing indicator while listening; the final text is appended to the current draft (never auto-sends);
- Clear error hints for denied permission / no microphone / restricted network / unsupported browser;
- Settings tab "Voice Input" (
settings.plugins.tab): browser support status + plugin load status;
- UI fully localized (zh / en) — follows the DSH language preference automatically.
Structure
src/
host/index.ts host half: /dsh-voice-input/health route (recognition itself is pure browser)
client/index.ts client half: mic button + Web Speech API + settings page (i18n zh/en)
scripts/
wrap-client.mjs wraps the client bundle with the official __ModuleLoader__.load protocol
self-test-client.mjs structure / load / registration assertions
tsup.config.ts dual build: host (esm, node) + client (cjs, browser)
cordis.patch.yml profile stack insertion patch
Develop / Build
npm install
npm run build # tsup build + client wrap + self-test
Outputs: lib/host/index.mjs (host), lib/client/index.js (client bundle).
Install into dsh
Local development (a file: reference; reinstall after code changes):
npm run build to make sure lib/ is up to date;- In the target profile's
package.json ($DSH_HOME/profiles/<profile>/package.json):
- add
"@hackerfish/dsh-voice-input": "file:<absolute path to the plugin dir>" to dependencies;
- add
"@hackerfish/dsh-voice-input" to dsh.profile.bundles;
- Run
pnpm install in that profile directory;
- Restart dsh and refresh the page — you will see the 🎤 button on the right of the composer.
Install from GitHub:
dsh plugin --profile web add https://github.com/hackerFish/dsh-voice-input
(or add github:hackerFish/dsh-voice-input to the profile's dependencies and run pnpm install.)
Notes & Limitations
- Recognition depends on the online speech service built into Chrome/Edge; when offline or unreachable, a "network restricted" hint is shown.
- The first click requests microphone permission (a
getUserMedia pre-check); a denial shows clear guidance.
- One utterance at a time: it stops automatically after a pause (
continuous=false), or click again to stop early.
- The transcript only fills the draft and never auto-sends, so misrecognitions can be corrected before sending.
