dshtools-sensevoice-input

Frequent voice-to-draft workflows in DSH Desktop often force you to ship audio to a cloud ASR endpoint, wait on round-trip latency, and accept a privacy trade-off you may not want. dshtools-sensevoice-input keeps the entire transcription pipeline inside the desktop process: tap the microphone next to the input box, record up to 60 seconds of speech, and let FunASR with SenseVoiceSmall run inference on a local CPU. The transcript is written straight into the draft via inputActions.setDraft, complete with parsed language and sentiment tags (e.g. zh · NEUTRAL), with zero outbound network calls.

The plugin lazy-loads the model on the first request—expect a 10-to-30-second warm-up—then reuses the same Python process for subsequent utterances of only a few seconds each. If the inference process crashes, it is restarted automatically, so you never have to relaunch the desktop app. The UI adapts to the DSH theme, a floating badge shows model readiness at a glance, and failures surface the exact reason with a one-tap retry.

It is built for DSH Desktop users who dictate often, treat audio as sensitive data, or work inside air-gapped and restricted networks. No cloud account, no API key, no audio ever leaves 127.0.0.1, and out of the box it covers transcription in more than 50 languages.