Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin,” which allows developers to extend host capabilities by mounting global plugins. In model inference scenarios, voice interaction is a key part of improving efficiency. moluyao/dsh-minimax-asr was created to address this pain point. It connects MiniMax’s speech recognition and synthesis capabilities directly into DSH, giving the model the ability to listen, speak, and interact automatically.

What It Is

This is a global plugin for DeepSeek Harness, designed to provide full voice interaction support for Harness instances. It integrates MiniMax’s asr-1.0 recognition model and speech-2.8-hd synthesis model, and uses global tools, settings cards, and UI components to turn complex API calls into simple operations. The plugin is open source under the MIT license.

Core Features

  • Global tools: Provides the transcribe_audio tool for transcribing audio files, and the announce_speech tool for speaking a voice announcement when a task finishes.
  • Interactive components:
    • Voice input: A voice input button is built into the input box, supporting recording and automatically appending the transcribed text to the draft.
    • Voice broadcast (TTS): Supports synthesizing text into speech, and can automatically read the result after a task finishes.
    • Real-time hands-free conversation: After enabling the “Conversation” button in the input box toolbar, the entire workflow can be operated hands-free: voice input is sent automatically, voice replies are announced automatically, and the microphone reopens automatically.
  • Voice selection: 303 built-in system voices are available for selection, supports grouping by language, and provides preview playback.

Installation and Enablement

Install it with the official command. The plugin is written to the profile as a bundle and takes effect after you restart dsh.

dsh plugin --profile web add github:moluyao/dsh-minimax-asr

After installation, you can find the corresponding settings card under “Settings → Plugins”. Configure the environment variables and it can be used.

Typical Usage

  • Automatic task announcement: After the model completes a task, it calls the announce_speech tool. The plugin automatically announces the result to the user through the browser speaker, so the user does not need to stare at the screen.
  • Hands-free interaction: After enabling conversation mode, the user only needs to speak. The plugin automatically transcribes and sends the input; after the model’s reply is announced, the microphone reopens automatically, with no need to touch the keyboard or mouse at any point.
  • Speech to text: Click the microphone button next to the input box. After recording ends, the audio is automatically transcribed and appended to the input box for easy secondary editing.

Applicable Scenarios and Notes

This plugin is suitable for scenarios that require voice interaction, hands-free operation, or voice broadcast capabilities.

  • File formats: Only formats such as wav, aiff, flac, m4a, opus, mp3, aac, and ogg are supported; webm and raw PCM are not supported.
  • File limits: The maximum size of a single file is 50 MB, or the duration is 500 seconds.
  • Key configuration: The API key is stored in the environment variable MINIMAX_API_KEY. The plugin resolves the key from that variable for every request and supports hot rotation.
  • Browser restrictions: Browsers require at least one prior user interaction with the page before allowing audio playback. Mic access must be granted on first use.
  • API parameters: response_format supports plain text, timestamped segments (verbose_json), and subtitle formats (srt/vtt).

Conclusion

moluyao/dsh-minimax-asr greatly simplifies the process of integrating voice capabilities into DeepSeek Harness by encapsulating global tools and UI components. Whether you need an assistant with automatic announcements or a scenario requiring hands-free interaction, this plugin provides a direct solution. For more details, see the GitHub repository or the DSH ecosystem directory.