dsh-voice-mode-desktop

English | 中文

Full-duplex voice mode for DeepSeek Harness 0.2 (desktop): speak and the words stream into the draft, pause and the message sends, the reply is read sentence by sentence, and speaking again interrupts the read-aloud. Ships an in-app settings page for reading speed, input mode and captions.

This is an adapted fork of dsh-voice-mode 0.7.15 (which in turn derives from haoku123/dsh-voice). Both are MIT licensed; see LICENSE.

What this fork is: the compatibility adaptation that makes that plugin run on DSH 0.2 desktop — it adapts the client API the plugin depends on, adds an in-app settings page, and replaces the cloud read-aloud dependency with a small client kept inside this repository. The recognition, barge-in and local-TTS engines are upstream work, unchanged.

• Upstream declares compatibility up to DSH 0.1.7-rc.1 and is written against the 0.1 client API.
• This fork targets DSH 0.2.0-rc.2 (the engine shipped with DSH NEXT desktop 2.0.17-next), where the client service settingsScope no longer exists and plugin settings forms only render volatile fields.
• Verified on Windows, DSH NEXT desktop 2.0.17-next, @deepseek-ai/* 0.2.0-rc.2. Other versions are untested.

What it does

• Voice in: press-and-hold or continuous listening, streaming recognition into an editable draft, silence-based auto-send.
• Voice out: sentence-by-sentence read-aloud (Edge cloud TTS by default; local VITS / Kokoro optional), live captions.
• Full duplex: speaking while the reply is read stops playback and cancels the running turn.
• Settings page: reading speed (0.5–2.0), input mode (hold / continuous), caption show-hide, and a one-click audio self-test. Reachable from the sidebar Plugins page as the item 语音模式 / Voice mode. The self-test deliberately separates synthesis from playback, so a silent reply can be traced to the engine or to the audio output instead of being guessed at.

Differences from upstream dsh-voice-mode 0.7.15

Area	Change	Why
Client half	inject no longer declares settingsScope; the access is wrapped in try/catch	settingsScope was removed from the DSH client API; the host treats "one entry did not activate" as fatal and the desktop app drops into recovery
Settings schema	rate and mode are marked volatile, and volatile values are unwrapped at every read site	DSH 0.2 renders only volatile fields in plugin forms; a wrapped value passed to the TTS engine makes synthesis fail silently
Settings UI	New in-app page (sidebar Plugins → 语音模式) plus host endpoint POST /voice-mode/settings; laid out like the host's own settings pages and themed with host design tokens (--dsw-alias-*)	The upstream settings card depends on the removed client API, and the host page that renders plugin config forms is not mounted in the desktop deployment
Captions	Added a show/hide switch; caption max width capped (420 px default tier)	Captions are optional; the previous default could span most of the window
Status bar	Width is content-sized, centered, capped at 480 px	It previously stretched across the whole composer
Diagnostics	One log line when a model stream is skipped because the voice session does not match	Makes a silent "no read-aloud" reproducible from logs
Packaging	Renamed package; the bundle patch mounts this package by a package-relative path	Makes the adaptation installable instead of a pile of local edits, and avoids the host's bare-specifier resolution for profile plugins
Cloud TTS	The msedge-tts dependency is replaced by an in-repo Edge read-aloud client (lib/edge-tts.js)	msedge-tts ships require("buffer/index"), which the DSH desktop host cannot resolve, so importing it failed and the plugin never loaded. Shipping the ~180-line client makes installation a single step with no dependency patching

Everything else — recognition, TTS engines, barge-in, wake word, backchannel yield, model pinning — is upstream code, unchanged.

Change scope

Baseline: dsh-voice-mode@0.7.15 as published on npm (itself derived from haoku123/dsh-voice). Files touched by this fork:

File	Nature of change
lib/index.js	host half: volatile fields in the settings schema, POST /voice-mode/settings, one diagnostic log line; the Edge engine now calls the in-repo client instead of msedge-tts
lib/client.js	client half: inject list, settingsScope guard, the settings page, caption show/hide and width, status-bar width
lib/edge-tts.js	new in this fork
cordis.patch.yml, package.json	mount point and packaging metadata (renamed package, no install scripts)
scripts/verify-compat.mjs, scripts/check-tts.mjs	new in this fork

lib/sense-worker.mjs and lib/tts-vits-worker.cjs are upstream files, byte-identical. The upstream project keeps its own git history, tests and build pipeline; this fork is a compatibility adaptation, not a rewrite.

Install

The plugin is a DSH bundle. In a profile:

# from the npm registry …
dsh plugin add @nutshelllee/dsh-voice-mode-desktop

# … or from a local checkout
dsh plugin add link:/path/to/dsh-voice-mode-desktop

# … or straight from the repository, without a registry entry
#   dependency: "@nutshelllee/dsh-voice-mode-desktop": "git+https://github.com/NutshellLee/dsh-voice-mode-desktop.git"

Then enable it in the profile manifest (dsh.profile.bundles) and, optionally, override defaults in the profile patch:

- id: voice-mode
  config:
    enabled: true
    voice: zh-CN-XiaoxiaoNeural
    rate: 1.1
    mode: hold          # hold = push to talk, toggle = continuous listening
    ttsEngine: edge     # edge | vits | kokoro
    bargeInMode: detect
    silenceMs: 1500
    autoSend: true
    spokenFormat: true

Restart the app after the first install, then open Plugins → 语音模式 to adjust settings, or press Ctrl+Shift+V inside a conversation.

Cloud read-aloud status

The Edge engine talks to Microsoft's read-aloud endpoint directly from lib/edge-tts.js (WebSocket handshake, SSML request, MP3 frames). No third-party TTS wrapper is involved, so installation needs no install script and no dependency patching. To check the path on your machine:

node scripts/check-tts.mjs

Verify

node scripts/verify-compat.mjs [path-to-dsh-app]

Checks the contract this fork depends on: the DSH app is present, all nine client inject targets exist, the host no longer serves settingsScope, the plugin's client half declares no missing service, and rate/mode project as live-editable fields. Exit code 0 means all checks passed.

Models

No model files are bundled. The plugin downloads what it needs the first time that capability is used, then reuses the local copy.

Capability	Downloaded when	Source repository
Streaming recognition (default)	first voice input	csukuangfj/sherpa-onnx-streaming-zipformer-zh-int8-2025-06-30 (encoder / decoder / joiner / tokens)
Voice activity detection	first voice input	Silero VAD (silero_vad.onnx)
SenseVoice recognition (senseVoice: true)	first voice input	SenseVoice int8 (model.int8.onnx + tokens)
Local VITS read-aloud (ttsEngine: vits)	first read-aloud with that engine	csukuangfj/sherpa-onnx-vits-zh-ll
Local Kokoro read-aloud (ttsEngine: kokoro)	first read-aloud with that engine	Kokoro int8 model set

• Hosts: huggingface.co first, hf-mirror.com as fallback. Only those two (plus hf.co) are accepted unless allowCustomHost: true is set in the profile patch. Set modelHost to pin a mirror.
• Cache directory: %LOCALAPPDATA%\dsh-voice-mode\models on Windows, ~/.cache/dsh-voice-mode/models elsewhere.
• Integrity: every file is checked against a pinned SHA-256 before use; partial downloads are written next to the target and only renamed once they verify.
• Progress: download and load progress is reported in the voice status bar; failures are surfaced in the chat rather than silently ignored.
• Offline: recognition needs the model files (and ttsEngine: edge needs network on every read-aloud). To run fully offline, pre-populate the cache directory and use vits / kokoro.

Known limitations

• Compiled artifacts: lib/ is upstream's built output with this fork's edits applied, not regenerated from source. Read the diff against dsh-voice-mode@0.7.15 rather than expecting readable source.
• Cloud TTS by default: ttsEngine: edge sends the spoken text to Microsoft. Use vits or kokoro for fully local synthesis.
• Barge-in on speakers depends on the browser's echo cancellation; on Windows, hold-to-talk or headphones give the most reliable interruption.
• Settings persistence covers rate and mode (the two volatile fields). Other options are read from the profile patch at startup.
• Diagnostics write to a local log: to pin down "an entire utterance went missing", the plugin writes every finalised segment (up to 400 characters) plus a set of pipeline events into DSH's application log (desktop-next.log). The log stays on the machine and the diagnostics endpoint listens on loopback only (POST /voice-mode/diag, same-origin checked); its content, however, is your own dictated text — clear the log periodically, and avoid installing this on a shared account.
• The correction dictionary is a local file: entries live in ~/.dsh/voice-dict.json (override with the DSH_VOICE_DICT environment variable) and are not shipped with this package; the repository carries the matching logic only, never anyone's entries.

Compatibility matrix

DSH (@deepseek-ai/*)	Desktop shell	Status
0.2.0-rc.2	DSH NEXT 2.0.17-next (Windows)	verified (load, settings, TTS, recognition)
0.1.7-rc.1 and older	—	use upstream dsh-voice-mode
other	—	untested

License and attribution

MIT. Copyright notices for the upstream projects are preserved in LICENSE. This fork is unofficial and not endorsed by the upstream authors; the package name is different on purpose.

The full adaptation log — what failed, in what order, and how each change was verified — is in docs/CHANGELOG.md.