dsh-dictation-hotkey
Press-to-talk dictation for the DSH Web composer. Hold a key combination, speak,
release — DSH's own speech engine transcribes it and the text lands in the
composer.
The plugin adds no second recognizer: it drives the speech Remote that DSH
already ships with its voice-input bundle (local SenseVoice, CPU) and inserts the
result through the official composer actions.
hold Ctrl+Shift+Space ──▶ speak ──▶ release
│
├─ microphone capture in the page (getUserMedia + AudioWorklet, muted output)
├─ encoded as the canonical 16 kHz mono PCM16 WAV the Remote requires
├─ ctx.remote.speech.transcribe() ← official speech Remote (SenseVoice)
└─ inputActions.insertText() / submit() ← official composer
What it does
- Push to talk:
hold (record while held) or toggle (press to start, press
to stop) trigger modes.
- Fill or send: insert the transcript for review, or submit it immediately.
- Reuses DSH's engine: the same local SenseVoice provider the official
microphone button uses — no extra process, model download, or cloud API.
- Composer-native insertion: text is spliced into the live draft through
inputActions.insertText, so the caret, references, and attachments survive.
- Configurable per browser: hotkey, mode, send behaviour, beeps, and
diagnostics live in a popover on the composer pill (
localStorage).
- Works on the new-conversation screen too: with no session open there is no
composer draft to splice into, so the transcript is pasted into the focused
composer as a normal edit, and the pill returns with the first session.
- Diagnosable: optional per-keystroke log with audio level, audio length, and
inference time, which separates "nothing was picked up" from "the recognizer
missed it".
Requirements
- DSH
0.2.0-rc.2 or a compatible release, Web profile.
- The official voice input bundle enabled (plugin manager → 语音输入). It
provides the speech engine; without it the pill reports that no recognizer is
mounted.
Install
# from a checkout
dsh plugin --profile web add /path/to/dsh-dictation-hotkey
# or straight from git (no clone, no build: the bundle is committed)
dsh plugin --profile web add github:AndyYang12345/dsh-dictation-hotkey
Then reload the Web UI. The bundle is committed at client/dist/client.js, so a
checkout works without a build step.
Updating later: dsh plugin --profile web add github:AndyYang12345/dsh-dictation-hotkey
again and restart dsh web — a profile keeps its own copy of the installed
package, so pushing to GitHub does not update an already-installed profile by
itself.
Use
A 🎙 Dictation pill appears under the composer:
| Action | Result |
|---|
Hold Ctrl+Shift+Space, speak, release | Transcript inserted into the composer |
| Click the pill | Settings: hotkey (press the keys to capture), fill ↔ send, hold ↔ toggle, beeps, diagnostics |
Releasing any key of the combination ends the recording, so there is no need
to lift them together.
On the home screen (no session yet) the pill is absent — the composer dock only
exists inside a session — but the hotkey still works: the controller stays armed
through a root-scope seat and the transcript is pasted into the focused composer.
Auto-send is skipped there, so sending stays your own keystroke.
Settings
Stored per browser under localStorage key dsh-dictation-hotkey.settings.
| Field | Default | Meaning |
|---|
hotkey | Ctrl+Shift+Space | +-separated combination; Space, F9, Alt+~, letters and digits are understood |
mode | hold | hold = record while held, toggle = press to start/stop |
autoSend | false | Submit the draft as soon as the transcript is inserted |
beep | true | Short tones for start, stop, and failure |
language | (empty) | Recognition language hint; empty follows the DSH speech settings |
providerId | (empty) | Speech provider id; empty uses the deployment default |
minRecordMs | 150 | Presses shorter than this are rejected with a visible notice |
tailMs | 180 | Audio kept after release, so a fast release does not clip a syllable |
maxRecordSeconds | 60 | Safety stop |
debug | false | Keep the keystroke log |
Diagnostics
Enable 诊断日志 in the popover; each keystroke then records a structured entry,
mirrored to the DevTools console and kept as a 20-entry ring buffer in
localStorage (dsh-dictation-hotkey.log, copy/clear buttons in the popover).
event | Meaning |
|---|
start | Microphone ready — micStartMs, triggering keys heldCodes |
stop | Released — heldMs, tailMs |
ok | Recognized — audioSeconds, inferenceMs, chars, peakDb |
empty-transcript | Engine returned nothing — carries peakDb; quiet: true means the recording itself was near-silent |
skip | Rejected by a threshold: shorter than minRecordMs, empty audio, speech remote absent |
transcribe-failed / mic-failed / insert-conflict / finish-threw | Corresponding failure |
peakDb > -45 with an empty transcript points at the recognizer (adjust
language); peakDb < -45 means that take carried no usable sound.
Development
npm run build # build client/dist/client.js (esbuild)
npm run verify # contract check: the bundle only touches injected services
npm test # behaviour suite: fake microphone + recognizer
verify.mjs exists because Cordis throws when a plugin reads a service it did
not declare (cannot get property "slots" without inject) while the Web shell
only prints web boot: 1 entry did not activate — a diagnostic dead end this
plugin hit once, now a build-time check. test-release.mjs covers sloppy release
order, double taps, sub-threshold presses, empty transcripts, and auto-send.
How it is wired
| Piece | Where |
|---|
| Hotkey, microphone, WAV encoding, state | client/src/{index,controller,hotkey,audio,settings,ui}.js |
| Composer seat (session-scoped) | conversation.composer.dock slot |
| Session-less seat | headless component in the root-scope shell.overlay slot |
| Recognition | ctx.remote.speech.transcribe — official speech Remote |
| Host half | host/index.js, intentionally a no-op (apply() {}) |
| Bundle layer | cordis.patch.yml inserts the plugin row |
The host half deliberately has no @deepseek-ai/* import and no service access:
a profile installs this package through a link, which resolves no dependencies for
it, so a host-side import would abort activation with ERR_MODULE_NOT_FOUND and
DSH would refuse to boot the Web GUI.
Limitations
- Keys are handled while the DSH page is focused. A key pressed with another
window in front never reaches the plugin; system-wide support needs an OS-level
helper and is not implemented.
- Microphone permission is granted for the DSH origin. The first ever use may
require clicking the official microphone button once.
- Recording is capped by
maxRecordSeconds and by the Remote's own audio limits.
License
MIT © 2026 AndyYang12345
