WizisCool/dsh-ears

https://github.com/user-attachments/assets/1363768e-a393-44bd-a008-1ce2055cac41

---

dsh-ears adds voice input and LLM-powered text polishing to DeepSeek Harness. It supports in-browser Web Speech, local Whisper, and popular speech-to-text (ASR) APIs.

Install

Prerequisites: DeepSeek Harness >=0.2.0-rc.1, and Node.js ^22.19.0 || >=24.0.0.

Install from npm

Web UI:

dsh plugin --profile web add dsh-ears

Desktop application: install from its plugin manager. The desktop app owns a reserved profile named desktop and the dsh CLI refuses to manage it, so there is no command-line equivalent. dsh-ears needs no separate desktop build — both surfaces load the same web client bundle.

Still using dsh 0.1.x? dsh-ears 0.4.3 requires dsh >=0.2.0-rc.1, and dsh refuses to install it on an older host. Stay on the matching earlier line:

# dsh 0.1.2 through 0.1.6
dsh plugin --profile web add "dsh-ears@<0.4.0"
# dsh 0.1.7 (dsh-ears 0.4.0 requires dsh 0.1.7-rc.2 or later)
dsh plugin --profile web add "dsh-ears@<0.4.1"
# dsh 0.1.1
dsh plugin --profile web add "dsh-ears@<0.3.0"

Install from source

git clone https://github.com/WizisCool/dsh-ears.git
cd dsh-ears
pnpm use:platform
pnpm install
pnpm build
dsh plugin --profile web add "$PWD"
# Windows cmd: use "%CD%"; PowerShell expands $PWD directly

pnpm use:platform switches node_modules to the native dependency tree for the active platform (kept separate for Windows and Linux). Run it on a fresh clone and whenever switching between platforms.

After installation, refresh the Web UI. A microphone icon appears to the right of the composer.

Update

dsh plugin --profile web add dsh-ears

add resolves the latest version from npm, so the same command works from any installed version. After updating, restart dsh web to load the new server code and refresh the Web UI; you can also check for new versions in the "About" panel on the settings page.

Uninstall

dsh plugin --profile web remove dsh-ears

Refresh the Web UI after uninstalling.

Usage

1. Click the microphone icon or press Ctrl+Shift+Space (configurable in settings).
2. Start speaking.
3. Press the shortcut again or click the microphone to stop recording and start transcription.
4. With polishing enabled, the raw transcript appears in the draft first. Once polishing completes, it replaces the text while preserving any manual edits made in the meantime.
5. Review the text and send manually.

Transcription results are always written to an editable draft and are never sent automatically. If the selected backend is not ready, the microphone icon is disabled — hover over it to see why.

Recognition backends

Backend	How it works	Requirements
Web Speech	In-browser real-time recognition	Default backend; Chromium-based browser
Local Whisper	Local transcription after recording stops	Download GGML models in settings
Groq	Groq Whisper API	API key
Deepgram	Pre-recorded audio or live audio streaming	API key, model name (e.g. nova-3)
Alibaba Cloud Model Studio	DashScope synchronous transcription	API key, synchronous model ID; max 300 s per recording
Tencent Cloud	Recording file recognition or real-time WebSocket	AppID, SecretID, SecretKey, engine_type
Xiaomi MiMo	Speech Recognition API, supporting standard API or Token Plan	API key, model name (e.g. mimo-v2.5-asr); Token Plan requires a regional cluster
SiliconFlow (CN)	Audio Transcription API	API key, model name (e.g. FunAudioLLM/SenseVoiceSmall)
Volcengine	Doubao audio file recognition or Doubao one-way streaming ASR	API key (new-console X-Api-Key), resource ID
Custom OpenAI-compatible	Sends to a specified /audio/transcriptions endpoint	Endpoint URL, API key, model name

Local Whisper: Runs whisper.cpp locally via @fugood/whisper.node without Python, FFmpeg, or external dependencies. Supports downloading and managing GGML models (tiny (default) through turbo) in settings, with default (auto), vulkan, or cuda acceleration.

Alibaba Cloud Model Studio: the upgraded console issues a workspace-specific API host (Beijing: https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com; Singapore: https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com). Enter that host and a synchronous model ID (qwen3-asr-flash, qwen-audio-3.1-asr-flash, or fun-asr-flash-2026-06-15), not the console display name and not a -filetrans, -streaming, -message, or -realtime variant — those use asynchronous or WebSocket APIs that dsh-ears does not support.

Volcengine: Accepts only the new-console API Key (X-Api-Key) authentication (legacy AppID + Access Token is not supported). Obtain keys from the console API Key page.

Polishing

Enabled by default. Leave both provider and model empty to use dsh's default Agent model (including its default reasoning settings); you can also select any model configured in dsh. LLM credentials reuse dsh's existing configuration. Reasoning effort is configurable.

The default prompt removes filler words, fixes common ASR errors, handles self-corrections, and formats lists. Customize the prompt or view the default content in settings. If polishing fails or is cancelled, the raw transcript is preserved.

Settings

After installation, a dsh-ears section appears in the Web UI settings page, organized into four tabs:

Tab	Configurable options
General	Settings display name, voice shortcut toggle and key combination, audio cues toggle, max recording duration (1–600 s, default 120 s)
Recognition	Backend selection, language, local Whisper model and acceleration, cloud providers and credentials
Polishing	Toggle, LLM provider and model, reasoning effort, custom prompt (up to 4,000 characters)
About	Version, license, dsh compatibility range, update checker

Known limitations

• The Web Speech backend relies on the browser implementation; audio may be sent to browser vendor servers for processing and is not a strictly local/offline solution.
• Alibaba Cloud Model Studio (Bailian) recordings are capped at 300 seconds per audio; Qwen-Audio 3.1 may return no words for very short or silent audio, so prefer qwen3-asr-flash for short voice input.
• Deepgram Flux models require the Listen V2 protocol and are currently unsupported.
• Local Whisper has a single-recording limit of 24 MB and a 120-second transcription timeout.
• Acceleration backend (default / vulkan / cuda) is locked after the first native module load; switching requires restarting dsh web.

Local development

pnpm use:platform
pnpm install
dsh plugin --profile web add "$PWD"
# Windows cmd: use "%CD%"; PowerShell expands $PWD directly
pnpm check          # type-check
pnpm test           # run tests
pnpm build          # build
pnpm dev:config     # build and write the HMR config
pnpm dev:web        # start dsh web

Run pnpm dev:watch in a second terminal while developing for live rebuilds.

After modifying frontend UI code, simply refresh the browser. When modifying server-side code, settings registration, Remote descriptors, or schemas, restart dsh web and then refresh.

Docs

• CHANGELOG
• CONTRIBUTING
• SECURITY
• LICENSE

License

MIT

Community Links

• LINUX DO — A new ideal community

Star History