dsh-voice-input

Typing is slow, and bouncing between a speech-to-text app and the editor to copy-paste results is a small but constant source of friction when working with DeepSeek Harness. dsh-voice-input collapses that into one gesture: a microphone button sits to the left of the input box, and whatever you say lands as text directly in the field. A live status bar floats above the input during recording, showing waveform level, a timer, real-time partial transcript, and Done or Cancel buttons, so you never leave the page.\n\nTwo engine paths are available. The zero-setup path uses the browser built-in SpeechRecognition API, ideal for quick dictation in Chrome or Edge. The higher-quality path speaks to any service that implements the OpenAI /audio/transcriptions contract: SiliconFlow, Groq, a local whisper.cpp server, or faster-whisper-server all work with a single preset click. The settings page includes a one-click local-backend panel that installs Whisper, starts Ollama, and pulls models through bundled PowerShell scripts, streaming logs back into the UI without ever opening a terminal. API keys are read on demand from the DSH credential store; nothing plaintext ever touches the config file, and rotating a key takes effect on the very next recording.\n\nThe plugin is honest about what does not work. After you select Ollama it inspects every local model for audio capability and, if none can listen, shows a red verdict with a one-click switch back to local Whisper. A registered voice_input_status tool lets you ask the model directly why voice input is not responding, and it reads the live configuration and points to the specific field to fix. This plugin is for people who frequently type long passages into DeepSeek Harness, users who want a fully offline speech-to-text pipeline, and anyone who would rather speak than type.