Introduction

When using the DeepSeek Harness Web interface, interaction with agents typically relies on text input. Existing voice plugins often only support Chinese and English, and may stop working when the browser loses focus. dsh-kitt-voice is a client plugin designed for DSH. Through an always-on-top floating window, it allows users to have natural voice conversations with agents and provides real-time feedback on operation status.

Plugin Positioning

The plugin is maintained by kittcat-lab and is licensed under the MIT license. It focuses on providing a conversational voice experience (rather than simple dictation) and integrates the Silero speech detector to effectively distinguish voice from environmental noise.

Core Features

  1. Voice Interaction and Detection
    * Conversational Voice: Press a key once to start speaking; the plugin automatically detects pauses, ends recording, and sends the input.
    * Noise Suppression: Uses the Silero detector to distinguish speech from noise, preventing false triggers in noisy environments.
    * Push-to-Talk: Supports talking by clicking the microphone or pressing a key; press again to send.

  2. Read-Aloud and Feedback
    * Real-Time Read-Aloud: When the agent replies, the content is read aloud sentence by sentence in real time, without waiting for the full response.
    * Interruption: Users can interrupt an ongoing response at any time.
    * Single-Voice Response: Each reply uses one fixed voice, engine, and speed, with no switching during the response.
    * Status Prompts: Provides real-time status reports (listening, transcribing, speaking), and plays an audible cue while listening.

  3. Global and Multi-Page Support
    * Global Hotkeys: Supports setting a global hotkey, allowing you to converse with the agent from any application.
    * Multi-Page Routing: When Harness is open on multiple pages, the global hotkey automatically routes to the page currently using voice or the current page.

  4. Voice Resources and Offline Support
    * Neural Voices: Provides 104 neural voices (45 Spanish, 47 English, 12 Chinese).
    * Piper Offline: Supports the Piper offline voice engine, allowing voice files to point to a local folder.
    * Device Selection: Supports selecting separate microphone and audio output devices.
    * Speed Adjustment: Supports adjustment from half speed to double speed, applicable to all engines (system voices, Piper, neural voices).

Installation and Enablement

Installing the plugin requires the official command. After installation, Harness must be fully restarted; simply reloading the page is not sufficient.

dsh plugin --profile web add dsh-kitt-voice

Typical Usage

  1. Key-Based Conversation: Click the microphone in the Composer toolbar, or press the preset shortcut key, then speak and press again to send.
  2. Global Control: Assign a global hotkey in the companion window or settings, enabling the agent even while using other applications.
  3. Offline Synthesis: Point to a local Piper voice folder in the plugin settings to perform fully local speech synthesis.
  4. Speed Control: Adjust the reading speed in the companion window menu without modifying configuration files.

Notes

  • Data Transmission: Reply text leaves the local machine and uses Microsoft’s read-aloud service (unless Piper is used).
  • Fault Handling: If the selected voice engine fails to read a reply, the remainder of that reply will use built-in voices, and the status bar will display a specific error.
  • Restart Requirement: After installing the plugin, Harness must be fully restarted, not merely reloaded.
  • English Phonemes: When using the Piper engine, English words (such as JSON) may be rewritten according to Spanish pronunciation rules; the plugin automatically corrects this.
  • Interface Language: The companion window UI supports Spanish, English, and Simplified Chinese, independent of the input language.