Introduction

DeepSeek Harness (DSH) aims to make the environment for developing and running agents fully web-based. For users who are accustomed to using mobile devices, the ability to directly control a DSH session on a computer via voice from their phone is a common need. The dsh-mobile-voice-call plugin fills this gap by combining a Web protocol with local speech engines.

Plugin Positioning

This is a DSH plugin that allows users to make voice calls to existing DeepSeek Harness sessions from a mobile browser. The plugin does not rely on a native mobile app. Instead, it reuses DSH’s Web Profile, transfers data over a Cloudflare tunnel, and runs Whisper.cpp and Hojo-TTS-Light locally for speech recognition and synthesis.

Core Features

  • App-free Mobile Voice Calls: No dedicated app needs to be installed on the phone; speech interaction happens directly through the browser.
  • Speech Recognition: Uses whisper.cpp to complete transcription locally, and recording data is not directly submitted to DeepSeek.
  • Speech Synthesis: Uses Hojo-TTS-Light to generate speech. If the engine is unavailable, it automatically falls back to Windows system speech.
  • Session Pairing: Securely connects the phone and DSH by scanning a QR code.

Installation and Enablement

Before installing, make sure that uv is available (required for running Hojo-TTS-Light).

pnpm dsh plugin --profile web add "github:SnowfallC/dsh-mobile-voice-call"
pnpm dsh web

Usage Steps

  1. Start the DSH Web Profile on the PC.
  2. Find and open the “Mobile Call” panel in the DSH interface.
  3. Scan the QR code on the screen using a mobile browser to pair the devices.
  4. On the mobile side, select an available top-level session.
  5. Allow microphone access, and you can begin the voice call.

Technical Details and Limitations

  • Environment Requirements: The page must be opened over HTTPS (Cloudflare Quick Tunnel is recommended), and uv must be installed.
  • Browser Compatibility: Built-in browsers such as those in WeChat and QQ are not supported; the page must be opened in a mobile browser.
  • Communication Mode: Uses a turn-based speaking mode, not a full-duplex voice call.
  • Data Transfer: Audio is transmitted over a Cloudflare tunnel and is not part of an end-to-end encrypted chain.
  • Interface Limitations: The mobile side provides only the voice call panel and does not include a text chat interface.
  • Speech Engines: On first use, the plugin automatically downloads a Whisper model of about 466 MB and a Hojo model of about 348 MB.

Summary

By combining local speech engines with Web transmission, this plugin enables voice control of DSH agents from a mobile phone without downloading an app. More details can be found in the GitHub repository.