@nibuil/dsh-voice-modes
为 DeepSeek Harness 编辑器提供语音输入，支持直接、润色、翻译和自定义角色模式，以及通过 URL 路由的 ASR 适配器。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @nibuil/dsh-voice-modes@0.2.0Voice Modes 以 @nibuil/dsh-voice-modes 发布，当前版本为 0.2.0。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。
Voice input for the DeepSeek Harness composer: a circular microphone button, a mode menu beside the model selector, and URL-routed ASR provider adapters.
Community plugin. Not an official DeepSeek Harness project.
Maintained by bin-r00t.
@nibuilis the npm publishing scope only; it does not refer to the GitHub usernibuil.
chat/completions + input_audio)/audio/transcriptions multipart)dsh plugin --profile web add @nibuil/dsh-voice-modes
Or install the bundle through DSH Plugin Manager. Restart the Web Host (dsh web) if DSH reports that a restart is required, then refresh the page.
After installation, click the mode dropdown in the composer and choose Voice settings.
| Field | Required | Notes |
|---|---|---|
| Endpoint | Yes | A /v1 base URL or a full provider endpoint. |
| API key | Yes | Stored on the Host only. |
| Model | Yes | Adapter-specific default is used when empty. |
The adapter is selected from the endpoint URL:
| URL contains | Adapter | Request |
|---|---|---|
mimo | Xiaomi MiMo ASR | JSON chat/completions with input_audio |
| anything else | OpenAI-compatible transcription | multipart /audio/transcriptions |
To force the MiMo adapter through a proxy URL, append a marker: https://proxy.example/asr#mimo.
Examples:
https://api.xiaomimimo.com/v1
https://api.xiaomimimo.com/v1/chat/completions
https://api.openai.com/v1
https://api.groq.com/openai/v1
Optional, but required for Polish, Translate and Custom role. It must expose an OpenAI-compatible chat/completions endpoint. The same provider is also tried as an ASR fallback when the primary ASR fails.
The settings dialog includes four examples: meeting notes, technical notes, business email and product request. Add, edit or remove roles freely.
English is the only target language in this release.
flowchart LR
A[Browser microphone] -->|WebM/MP4| B[16 kHz mono WAV]
B -->|POST /voice-modes/process| C[DSH Host proxy]
C --> D{URL router}
D -->|contains mimo| E[Xiaomi MiMo adapter]
D -->|fallback| F[OpenAI transcription adapter]
E --> G[ASR transcript]
F --> G
G -->|Direct| H[Composer draft]
G -->|Polish / Translate / Role| I[Second provider]
I --> H
The browser never calls an ASR provider directly and never receives a stored API key. The Host stores settings in ~/.dsh/voice-modes/settings.json with mode 0600 and exposes same-origin routes protected by DSH's connection authentication.
See docs/architecture.md for details.
Requirements: Node.js 20+ and a DSH installation.
git clone https://github.com/bin-r00t/dsh-voice-modes.git
cd dsh-voice-modes
npm test
npm run check
Install the working copy through DSH Plugin Manager or:
dsh plugin --profile web add "file:$PWD"
Restart dsh web after Host-side changes. Client-only changes reload through the Web HMR channel; refresh the page if they do not.
Test the whole flow without a real ASR account:
node debug/mock-provider.mjs
Use http://127.0.0.1:8787/v1 for the generic adapter, or http://127.0.0.1:8787/v1#mimo for the MiMo adapter.
Read docs/adding-asr-provider.md, add the adapter to providers.js, and extend the tests.
0.2.1-rc.2.engines.dsh; incompatible DSH versions are rejected before the plugin is imported.conversation.input.activity, ctx.slots, and ctx.locale. These are current DSH extension points, not a frozen public API.Contributions are welcome, especially new ASR adapters, browser compatibility fixes, translations and tests.