dsh-voice-ptt

English | 中文

Push-to-talk dictation for DeepSeek Harness: hold a single key in the composer, speak, release — the transcript is inserted into your draft.

No chord to memorise. Hold F2, release, done.

---

What it does

1. Put the cursor in the composer. While it is focused and the draft is empty, a faint hint appears in the toolbar: Hold F2 to dictate.
2. Hold F2 — recording starts and the hint becomes Recording — release F2 to finish.
3. Release — the audio is transcribed and the text lands at the caret in your draft.

While dictating:

• Press Esc, or press any other key (say, the letter you meant to type), to cancel and discard the recording.
• Losing window focus, or switching the tab away, cancels the recording rather than capturing unrelated audio.
• Exceeding the recognizer's limit stops the recording automatically and transcribes what it got.

Changing the key

The hint is a button. Click it to expand the candidate keys, click one, and it applies immediately and is remembered (in localStorage).

Key	Notes
F2 (default)	Function key: types no character, minimal conflicts
F4	Function key, further right
Insert	Right hand on full-size keyboards; absent on many laptops
Right Ctrl	Comfortable for the right pinky
Right Shift	Right pinky; on Windows, five taps can trigger Sticky Keys

Modifier keys work properly: the plugin knows the hold key is a modifier, so it does not mistake it for a chord and pass the event through.

The hint, the status line, and these key names follow your DSH language (en / zh).

To offer a different key, edit KEY_OPTIONS at the top of lib/client.js (a {code, labelKey} array, code spelled as KeyboardEvent.code, labelKey naming an entry in the zh / en dictionaries at the top of the same file).

Install

dsh plugin --profile web add dsh-voice-ptt

The package declares dsh.bundle.patch, so the command appends it to the profile's bundle stack and mounts it — no hand-editing of cordis.patch.yml. From a Git checkout, pin the revision:

dsh plugin --profile web add 'github:eicon-xyz/dsh-voice-ptt#<commit>'

Requires the speech recognizer. Install and enable @deepseek-ai/dsh-experimental-voice-input-bundle and prepare a model once (roughly 230 MB). If the recognizer is not ready, holding the key reports a visible reason in the status line instead of failing silently.

Development

# mount the checkout directly instead of installing
ln -sfn "$PWD" "$DSH_HOME/profiles/web/node_modules/dsh-voice-ptt"
# then add to the profile's own cordis.patch.yml:
#   - insert:
#       - id: dsh-voice-ptt
#         name: 'dsh-voice-ptt'

Remove that manual mount line before switching to the bundle channel, or the row is mounted twice.

Why a single key needs its own keyboard handling

The official shortcut service (@deepseek-ai/dsh-client-shortcuts) dispatches keydown only — there is no keyup event to end a hold. It also rejects modifier-free bindings (bindingIssue returns modifier-required) and, on the web, restricts chords to a platform allow-list.

So "hold one key and talk" cannot be expressed as a shortcut command. This plugin owns the keyboard instead: document-level keydown starts, keyup finishes. The trade-off is that it cannot appear in Settings → Shortcuts, which is why the key picker lives in the composer.

Keeping a single key from becoming a second modifier

Situation	Behaviour
Focus is not in a composer	Does not start recording
Any other key pressed while recording	Cancels the take; that key works normally (hold Right Ctrl, press A → Ctrl+A still applies)
A dialog or menu is open	Yields the key
Modified press (Ctrl+F2, Alt+F2, …)	Not intercepted at all
Focus is in a terminal	Not intercepted (xterm needs function keys)

How it works

• Draft insertion uses the session-scoped inputActions (captureInsertion + insertText). The plugin registers a resident entry in conversation.input.left to record each session's actions; that same entry renders the status line, the focus hint, and the key picker.
• Recording uses MediaRecorder and resamples through OfflineAudioContext to 16 kHz mono PCM16 WAV — the same format the official voice-input plugin sends — then calls the host's speech.transcribe.
• The host half is deliberately empty. Every behaviour lives in the browser half, reached through dsh.client.

Files

Path	Purpose
lib/client.js	Browser half: keys, recording, transcription, status line, key picker
lib/index.js	Host half (intentionally empty, so the bundle row is mountable)
cordis.patch.yml	dsh.bundle.patch layer the installer merges
test/	Headless-Chrome end-to-end tests

Tests

They drive a real browser against a running DSH web profile, with Chrome's fake microphone fed the bundled fixture:

npm install                       # provides ws
export PTT_TOKEN=<token from the dsh web URL>
node test/e2e.cjs                 # hold → release → transcript inserted
node test/picker.cjs              # picker opens, a new key applies and persists
node test/modifier-key.cjs        # Right Ctrl records; Ctrl+A is not hijacked
node test/nonempty-draft.cjs      # works with a non-empty draft, text is appended
node test/modal-yield.cjs         # yields while a dialog is open, works again after

test/make-fixture.cjs regenerates test/fixtures/speech-16k.wav from any audio Chrome can decode:

node test/make-fixture.cjs input.mp3        # needs Chrome; no ffmpeg required

Limitations

• Web surface only (the desktop native keyboard bridge is not wired to this plugin).
• No streaming captions: the text appears after you release.
• Never sends automatically — the transcript goes to the draft for you to review.
• One hold key at a time, and only from the built-in candidate list unless you edit KEY_OPTIONS.

License

MIT