Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture, with its ecosystem contributed by the community. dsh-voice-call is a personal voice-call assistant plugin designed to fill the gap in DSH’s native voice interaction. It automatically creates a dedicated workspace and session to provide persistent conversations across restarts; provides a floating-ball panel for quick access; uses FunASR for speech recognition and cloud TTS for speech synthesis; and dispatches sub-agent tasks to independent sessions via the voice_task tool while tracking progress.

Core Features

  • Dedicated Workspace and Session: After installing and restarting DSH, the Host automatically creates a dedicated workspace “Voice Call” and a dedicated session voice-call-main. This session persists across DSH restarts and does not need to be created manually.
  • Floating Ball Call Panel: A draggable floating ball appears in the lower-right corner of the page. Click it to open the Call, Tasks, and Settings panels. It reappears automatically after restart.
  • Speech Recognition: Supports FunASR HTTP mode. The default URL is http://127.0.0.1:10095/v1/audio/transcriptions, and the model is fun-asr-nano. Supports streaming and cloud APIs.
  • Voice Response: Supports MiniMax, MiMo-V2.5-TTS, and OpenAI-compatible TTS. Supports speech-rate adjustment and maximum-character truncation detection.
  • MiMo Director Mode: In MiMo-V2.5-TTS mode, director mode is supported (Off / Fixed Script / Auto Each Turn). Voice cloning (VoiceClone) is supported without registering a voice_id; samples are sent as DataURLs.
  • Task Dispatch: The main session dispatches tasks to independent task sessions via the voice_task tool or the panel. After a task completes, the result is automatically read aloud.
  • Wake Word Mode: The wake word (default: “Little Whale”) and idle timeout (default: 8 seconds) are configurable. Using local ASR is recommended to reduce latency.

Installation and Enabling

This plugin provides a persistent installation method without manually editing configuration files. The installation command is:

dsh plugin --profile web add @biliye/dsh-voice-call

After installation, restart DSH. Under Settings → Plugins, dsh-voice-call should be visible. After restart, the Host automatically creates the dedicated workspace “Voice Call” and the session voice-call-main.

Configuration and Usage

  1. Configure Speech Recognition: In the Settings page opened from the floating ball, configure the speech recognition engine. For FunASR mode, enter http://127.0.0.1:10095/v1/audio/transcriptions as the URL and fun-asr-nano as the model.
  2. Start a Call: Click ▶ Start Call on the floating ball. Speak directly and pause for about 1.2 seconds; the system automatically recognizes and sends your speech. After the assistant replies, it automatically reads the reply aloud.
  3. Dispatch a Task: On the Tasks page or by saying “Help me check…”, the system dispatches the task to a sub-agent for execution. After the task completes, the result is automatically read aloud.
  4. Use MiMo Director Mode:
    • In Voice Response (TTS) on the Settings page, switch TTS Model to MiMo-V2.5-TTS.
    • Enter the API Key (obtained from the platform.xiaomimimo.com console; no GroupId is required).
    • Select a voice (built-in: Rock Candy / Jasmine / Soda / Birch / Mia / Chloe / Milo / Dean).
    • Enable Director Mode, and choose Off, Fixed Script, or Auto Each Turn.

Notes

  • Local ASR Recommendation: For privacy and more reliable wake-up, it is recommended to use a locally deployed FunASR Server.
  • Voice Cloning Guidelines: ⚠ Only clone your own voice or a voice for which you have obtained explicit authorization. Samples support mp3/wav files, with a size limit of 10 MB.
  • Development Environment Configuration: If you develop the source code locally and install it using dsh plugin add ., the default is a copy-based installation due to the nodeLinker: hoisted configuration. Source code changes require a restart to take effect. If hot reloading is needed, replace the installation directory with a Junction, but note that the source directory must include its own node_modules; otherwise, host dependency resolution may fail.

Ecosystem Background

The DSH philosophy is “everything is a plugin.” The community directory skillhub.cn provides plugin cataloging and search. This document is written based on verified facts. For specific source code and updates, refer to the GitHub repository.