dsh-voice

DeepSeek Harness web GUI is purely text-based, and reading through long or code-heavy responses is tiring and slow. dsh-voice adds voice messaging to the conversation so that agent output can actually be heard.

The plugin works on two sides. On the Node side it registers a voice_say tool that the agent can call to send voice messages on its own, and exposes an HTTP TTS endpoint. Before synthesis, text is cleaned to strip fenced and inline code, URLs, HTML tags, table markup, and headings; the result is handed to Microsoft Edge TTS with no API key required and cached locally to deduplicate parallel requests. In the browser, a Listen button and inline audio players appear at the tail of each agent turn, with a male and female voice toggle persisted in localStorage. Only one player plays at a time, backed by a 30-entry blob cache.

If you read agent output on a phone, in a car, or simply prefer listening over staring at a screen, dsh-voice is a zero-config audio layer. Russian male and female voices are the defaults; any other language can be selected by passing a full Edge TTS ShortName. The only requirement is network access to speech.platform.bing.com.