dsh-voice-kit

The DeepSeek Harness web GUI has always been a purely text-based experience: type to talk, read to understand. Long sessions with the model can tire your hands, and lengthy replies are easy to skim without absorbing. dsh-voice-kit brings real voice interaction into DSH. A microphone button appears in the composer, and dictated speech is transcribed on the host side and appended to your draft without ever overwriting existing text. Every assistant message gains a read-aloud button at its tail for one-tap listening.

Read-aloud offers three engine tiers for different network conditions. Microsoft neural voices are free and natural, synthesized locally via msedge-tts with caching, and automatically routed through your system proxy on restricted networks. Cloud neural voices connect to SiliconFlow CosyVoice2 or any OpenAI-compatible speech service, work directly from mainland China without a proxy, and share the same credential you configured for voice recognition. System voices provide a fully offline fallback that works everywhere. Playback streams in sentence-boundary chunks: the first chunk is audible within roughly one second while later chunks download in the background. Markdown formatting and emoji are stripped before speaking, the viewport scrolls to the active message, and an on-screen bubble shows what is being read.

The plugin is well suited for DSH Desktop users who want to reduce typing fatigue, consume long AI replies hands-free, or keep a smooth voice workflow even behind proxies. Voice recognition supports two paths: the browser Web Speech API or a host-side OpenAI-compatible transcription service, so at least one will work in most setups. The settings panel includes one-tap presets, a live reachability probe for each engine, and a key-status indicator, making troubleshooting far less guesswork.