dsh-voice-input

Speech transcriptions often arrive cluttered with filler words, stammered repetitions, and missing punctuation, forcing a manual tidy-up before they are usable in DeepSeek Harness. dsh-voice-input fills that gap: it adds a one-click microphone to the Harness conversation composer, uses the model already selected in your current session for a tightly constrained cleanup pass, and inserts the polished result straight into the input field. No extra API keys, no leaving Harness.

It recognises Mandarin, Traditional Chinese, Cantonese, English, Japanese, and Korean with automatic language detection. It prefers the browser SpeechRecognition API and falls back to OS dictation (Fn twice on macOS, Win+H on Windows) when browser support is absent. The default hotkey is Ctrl+Alt+V, rebindable to any modified shortcut or F1-F24 while the Harness page has focus. The cleanup contract is deliberately strict: remove meaningless fillers, speech-error repetitions, and explicit self-corrections; add sentence punctuation and paragraph breaks; use Markdown only when the speaker clearly dictates a list or heading. It does not translate or answer dictated content. If cleanup fails, the raw transcript is retained and any edits made while waiting are preserved.

It suits users who primarily type prompts in DeepSeek Harness but occasionally want to dictate in a mixed Chinese-English context without opening a separate transcription tool.