dsh-voice

The DeepSeek Harness Web GUI has lacked a native voice-input path, forcing reliance on the keyboard even for lengthy prompts. dsh-voice adds a microphone icon beside the send button; tapping it reveals a ChatGPT-style recording pill with three actions — cancel, transcribe-and-insert, or transcribe-and-send — making voice as natural a way to compose as typing, all within the same browser tab.

Under the hood, the plugin ships three switchable transcription engines. The default is transcribe.cpp, which carries a 67-model catalog (GigaAM v3 CTC/RNN-T/E2E, Voxtral Mini 3B/4B/24B, Whisper tiny through large-v3 + turbo, Qwen3-ASR, SenseVoice, and more) with one-click Hugging Face downloads and progress bars. For fully offline or air-gapped setups, whisper.cpp with any GGML model works out of the box. Enterprise or cloud deployments can point the plugin at any OpenAI-compatible /v1/audio/transcriptions endpoint. Recordings longer than 22 s are automatically segmented into 20 s chunks, batch-transcribed in a single model load, and stitched back together so no audio is silently dropped. Live waveform feedback scrolls in real time with history fading into a dotted baseline.

Ideal for DSH users who type a lot and would benefit from hands-free input: developers wanting high-accuracy ASR without sending audio to a third party, enterprise teams already running private or cloud transcription services, and anyone who simply prefers speaking to their harness the way they speak to a phone.