Foreword

The web interface of DSH is typically text-based. Keyboard input isn’t always ideal when supplementing commands, processing natural language, or quickly recording content via voice. dsh-ears is an open-source voice input plugin for DeepSeek Harness that integrates speech recognition, transcription, and optional refinement into the DSH Web UI. It does not automatically send messages for the user; instead, it converts speech into an editable draft.

What Is This

First, let’s clarify the target. dsh-ears is a voice input plugin for the DSH Web interface, hosted at:

WizisCool/dsh-ears

The license is MIT. The version recorded in package.json is:

0.2.1

It primarily handles three tasks: selecting a speech recognition backend, writing transcription results to a draft, and optionally refining the transcribed text with a polishing model.

Core Features

Recognition Backends

Here are the supported recognition backends:

Web Speech
Local Whisper
Groq
Deepgram
Alibaba Cloud Bailian
Tencent Cloud
Xiaomi MiMo
SiliconFlow
Custom OpenAI-compatible recognition backend

The local Whisper backend requires downloading a GGML model on the settings page. Additional requirements for some cloud backends are as follows:

Alibaba Cloud Bailian: Maximum 300 seconds per request
Tencent Cloud: AppID, SecretID, SecretKey, engine_type
Xiaomi MiMo: Token Plan requires selecting a regional cluster

Refinement and Drafts

Refinement is enabled by default. The model can use the default DSH Agent model or be selected from DSH’s configured models; the prompt can be customized.

Transcription results are always written to an editable draft and are not automatically sent. When refinement is enabled, the original transcription is first written to the draft, and then updated after refinement is complete.

Installation and Updates

Prerequisites

Before installation, confirm the version range:

DeepSeek Harness: 0.1.0-rc.6 to 0.1.1-rc.2
Node.js: ^22.19.0 || >=24.0.0

Installation

First, add the plugin:

dsh plugin --profile web add dsh-ears

Updating

Use the same add command when updating:

dsh plugin --profile web add dsh-ears

Then restart dsh web and refresh the Web UI to load the new host code.

Uninstallation

dsh plugin --profile web remove dsh-ears

Typical Usage

After installation, the usage flow is as follows:

  1. Click the microphone icon, or press:
Ctrl+Shift+Space

to start speaking.

  1. Press the same shortcut again, or click the microphone icon, to stop recording and begin transcription.

  2. If you choose the local Whisper backend, first download the GGML model on the settings page.

  3. When refinement is enabled, the original transcription is first written to the draft, and then updated after refinement is complete.

  4. After reviewing the draft content, send it manually.

Applicable Scenarios and Notes

Suitable scenarios:

  • You are already using DeepSeek Harness in the DSH Web environment and want to input commands or supplement context via voice.
  • You want the transcription results to enter a draft for manual review before sending.
  • You need to choose between Web Speech, local Whisper, cloud recognition backends, or a custom OpenAI-compatible backend.

Please note:

  • The plugin runs with the permissions of the current dsh process; it is recommended to review the source code, dependencies, and MIT license before installation.
  • All API keys and credentials are stored by the host and are not accessed by the browser.
  • The version compatibility range is:
DeepSeek Harness: 0.1.0-rc.6 to 0.1.1-rc.2
Node.js: ^22.19.0 || >=24.0.0
  • Transcription results are not automatically sent; final confirmation is still required manually.

Conclusion

dsh-ears implements voice input as a DSH Web plugin: first select a recognition backend, then place the transcription results in a draft, and finally allow the user to send them manually.

Repository URL:

https://github.com/WizisCool/dsh-ears