@nutshelllee/dsh-voice-mode-desktop
适用于 DeepSeek Harness 0.2 桌面版的全双工语音模式：将流式 ASR 转换为草稿、句子朗读、插话、实时字幕，以及应用内设置页面。dsh-voice-mode 0.7.15 的非官方兼容分支。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @nutshelllee/dsh-voice-mode-desktop@0.1.5Voice Mode Desktop 以 @nutshelllee/dsh-voice-mode-desktop 发布，当前版本为 0.1.5。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
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.
2.0.17-next), where the client service settingsScope no longer exists and plugin settings forms only render volatile fields.2.0.17-next, @deepseek-ai/* 0.2.0-rc.2. Other versions are untested.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.
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.
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.
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
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.
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 |
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.%LOCALAPPDATA%\dsh-voice-mode\models on Windows, ~/.cache/dsh-voice-mode/models elsewhere.ttsEngine: edge needs network on every read-aloud). To run fully offline, pre-populate the cache directory and use vits / kokoro.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.ttsEngine: edge sends the spoken text to Microsoft. Use vits or kokoro for fully local synthesis.rate and mode (the two volatile fields). Other options are read from the profile patch at startup.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.~/.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.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 |
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.