Introduction

DeepSeek Harness (DSH) relies on a plugin-based architecture and can connect various capabilities to agent applications through plugins. Text-to-speech is an important part of improving the interaction experience. dsh-gsv-tts is a text-to-speech plugin built for DSH, supporting two modes: Edge cloud zero-configuration quick experience and GSV local offline voice cloning.

Core Features

This plugin primarily addresses speech synthesis needs in DSH environments and provides the following capabilities:

  1. Dual-mode speech synthesis

    • Edge cloud mode: Zero configuration, ready to use after installation. It uses Microsoft’s curated voices (20+ options), can read aloud when online, and typically responds within 1 second.
    • GSV local mode: Fully offline. It supports cloning a target voice from a reference audio sample, suitable for privacy-sensitive or network-less scenarios. The engine takes about 15 to 90 seconds to load the model on first startup.

  2. Interaction controls

    • One-click read-aloud: A 🔊 button appears next to each assistant message; click it to read the message aloud. The feature automatically excludes “thinking” content and reads only the response text.
    • Automatic read-aloud: When enabled, assistant replies are read aloud automatically. New replies interrupt the current speech by default (barge-in) to prevent overlapping audio; this interruption can also be disabled in the settings.

  3. Management and configuration

    • Mode switching: The settings panel supports one-click switching between “Quick Experience” and “Local Professional” modes, with changes taking effect in real time.
    • Voice settings panel: Provides a visual configuration interface, with settings applied hot after saving. Supports previewing individual voices or all voices.
    • Voice marketplace: Supports one-click download and injection of voices from built-in manifests or custom remote manifests. Built-in trusted sources do not require secondary confirmation; remote sources require confirmation of provenance and license.

  4. Engine and resource management

    • One-click engine start/stop: The local GSV engine supports one-click start and stop.
    • One-click engine installation: Provides automatic installation scripts, detects the Python environment, and deploys the GSV-TTS-Lite engine.
    • Same-origin audio short links: Synthesized audio is saved to disk and returned as a same-origin short link, avoiding bloating the model context.

Installation and Enablement

Use DSH’s command-line tool to install the plugin. The following command uses the web profile as an example; desktop applications usually use the desktop profile, so replace web in the command.

dsh plugin --profile web add dsh-gsv-tts

After installation, restart DSH. The plugin tools will be automatically registered with the Agent, and a “Voice Settings” option will be added in settings.

Typical Usage

Quick Experience Mode (Edge)

  1. Install DSH and restart it.
  2. Open Settings → Voice Settings.
  3. Select “Quick Experience (Edge)” for TTS mode.
  4. Choose a cloud voice from the voice dropdown (e.g., Xiaoxiao, Yunxi, etc.).
  5. Click the 🔊 button next to any assistant message to hear the voice.

Local Professional Mode (GSV)

  1. Install DSH and restart it.
  2. Open Settings → Voice Settings and switch TTS mode to “Local Professional (GSV)”.
  3. Turn on the “Engine Switch” in the settings panel and wait until the status shows “Running”.
  4. Add a voice preset or download a voice from the “Voice Marketplace”.
  5. Click the 🔊 button next to a message, or let the Agent call the tts_speak tool to read aloud.

Notes

  1. Network dependency: Edge mode requires an internet connection. If the cloud service is unavailable, the panel prompts you and guides you to switch to GSV mode.
  2. Local environment: GSV local mode requires a Python 3.10+ environment. The model is downloaded on first startup, and loading takes about 15 to 90 seconds.
  3. Permissions and security: Installing remote manifests from the Voice Marketplace requires secondary confirmation of provenance and license. Marketplace-installed voices are read-only in the panel (only uninstall allowed).
  4. Configuration fields: The quotaDaily field in the configuration is a reserved guidance field (default null means unlimited); hard quotas are planned for future releases.
  5. Source review: The plugin is open source under the MIT license. It is recommended to review the source code and license before installation to ensure they meet the environment’s security requirements.