dsh-voice-input

Voice input has been a missing piece in the DeepSeek Harness web chat for a long time—every thought had to be typed out one character at a time. dsh-voice-input fills that gap by placing a microphone button right beside the input field: tap once to start recording, tap again to stop, and the transcribed text lands in the box automatically, with no window-switching or external tools required.

All transcription runs on a local SenseVoice service, so no audio ever leaves your machine. The first time you hit the mic, the plugin silently downloads the roughly 230 MB model while a ring animation and a live percentage counter appear on the button itself. A separate gear button lets you inspect or remove the installed model whenever you like. Python dependencies (flask, onnxruntime, librosa, and friends) and the backend processes are also handled automatically when dsh starts up; the only thing you need to prepare yourself is a Python 3.10+ interpreter on the system PATH.

It is built for writers and developers who think in voice and want to cut down on long typing sessions, as well as researchers who prefer to keep audio data entirely on-device. If your DeepSeek Harness runs on the Web profile and your browser supports the MediaRecorder API, you are all set to plug in and go.