Say the task. Keep the final word.

Use the microphone in a new task or an existing conversation. Stop recording to turn your speech into text at the saved cursor position. Nothing is sent automatically: you can correct a name, add a detail or change the whole request before starting work.

If you keep typing while recognition finishes, VECTA preserves those edits and offers the recognised text for you to insert. Changing tasks or cancelling discards the recording instead of adding a late answer to another draft.

First use explains what will be downloaded. Whisper Small is the default: about 488 MB (465 MiB). The optional Base model is about 148 MB (141 MiB). Downloads show progress and can be cancelled; downloading never starts the microphone.

Choose the speech language independently of your coding agent, or use automatic detection. Voice settings include microphone selection, an input-level check, model selection and removal. A new model does not replace your selected model until you choose it.

Recognition uses Whisper through whisper.cpp on your Mac. There is no hosted transcription fallback or per-request transcription charge. Downloading a model needs an internet connection; using the installed model does not.

Audio and temporary recognition files are removed after processing or cancellation. The resulting words are still a draft. When you press Send, that text becomes part of the task and follows the agent or provider connection you selected.

The writing area brings agent and run controls into the composer instead of leaving them in a separate row. Starting suggestions stay close to the draft, so you can choose an opening and refine it before you send.

Read-aloud is off by default. Enable it in Voice settings, choose a local system voice, then use Read answer aloud beside Copy on a completed agent reply. Stop reading ends playback; new answers are never read automatically.

Only voices reported as local by the browser or app are offered. Availability depends on your device and browser, in both desktop and web. There is no hosted voice fallback, and reading an answer does not open the microphone.

The microphone menu in an existing task also offers Voice command. Say “Show changes”, “What is the agent doing?” or “Stop task”, then review the recognised phrase and confirm the proposed action.

Commands use the task’s existing panels and permissions. Recognition never executes an action by itself. Unknown phrases can stay as an editable message; publishing, deleting and approving are not voice commands.

Local dictation starts in the VECTA desktop app for Macs with Apple Silicon. It is not available in the web app, on Intel Macs, Windows or Linux in this release. Update the desktop app through your approved account to get the new composer and voice controls.

Speech recognition can make mistakes, especially with names, numbers and mixed-language phrases. Read the draft before sending. The documentation explains setup, model choices and how to recover when the microphone or download is unavailable.