Live capture from the AgentRQ app — the mic button's entry point into on-device dictation.
Speech never leaves the tab it was recorded in.
Every description and reply field has a mic button next to it. Click it, grant microphone access once, and start talking.
A Whisper model downloads once (with visible progress) and then runs in a background Web Worker — transcription doesn't block the rest of the page.
The transcript is inserted directly into whatever field you were dictating into — keep editing it like anything you'd typed yourself.
There's no speech-to-text API call here — no audio blob uploaded to a transcription service, no third party ever hearing what you said. The model runs in your browser, on your hardware, and the only thing that leaves the Web Worker is the text it produced.
That matters for the same reason auto-title generation runs locally too: task descriptions and replies often contain internal hostnames, file paths, or context you wouldn't want leaving your machine just to get transcribed.
Whisper is multilingual by default — AgentRQ ships a language picker
(whisperLanguages.js) so you can
dictate in the language you actually think in, not just English.
Like auto-title generation, the model is downloaded once and cached by the browser — every dictation after the first is instant, with no repeat download.