dsh-voice

Typing long prompts or code snippets into DeepSeek Harness slows you down. dsh-voice plugs local, private speech-to-text directly into the workflow: a microphone button appears beside the composer in the Web UI, and the TUI can capture audio through a system recorder. Speak, and the transcript lands where your cursor is—no tab switching, no cloud round-trip.

Under the hood the plugin runs a fully offline ONNX Whisper pipeline. The default model (whisper-medium, q8, roughly 0.8 GB) is fetched from the Hugging Face Hub on first use, cached locally, and inferred via onnxruntime-native with no external binaries and no network calls at runtime. It exposes both a ctx.stt service for programmatic use and two HTTP routes (/voice/transcribe, /voice/status) so other plugins or scripts can share the same transcription path. Need a lighter 250 MB small model or the higher-accuracy 1.6 GB large-v3? A single line in cordis.patch.yml switches it, and the old pipeline is released once no transcription is still holding it.

It is built for people who already live inside dsh—writing prompts, steering agent loops, or quickly capturing ideas and code in the terminal. The Web side requires nothing beyond the browser; the TUI side just needs a system recorder such as ffmpeg or sox. Once the model is cached, every subsequent transcription is instant and stays on your machine.