1. Use Cursor’s native voice input first
Update Cursor, open the agent input where voice is available, and press and hold Control-M. Cursor’s current interface shows a waveform, timer, and controls to cancel or confirm the recording.
Native input is the shortest path when the destination is already Cursor. It owns the input surface and keeps the prompt attached to the active agent context.
2. Treat batch speech-to-text as dictation, not live conversation
Cursor says the feature records the full voice clip and transcribes it with batch speech-to-text. That gives you text for the agent; it is not the same interaction model as a continuous spoken conversation with audio replies.
Review the transcript and the repository target before confirming, especially when the prompt can start an agent action.
3. Check current data controls instead of inferring
Cursor’s voice changelog explains the interface but does not specify the voice audio processor or retention rule. Do not infer local or cloud processing from the words “batch STT.”
Cursor’s Data Use page explains Privacy Mode, zero-data-retention agreements, risk-classifier exceptions, codebase indexing, and temporary caching for the broader service. Review that current page and your workspace settings before sensitive voice use.
4. Use IraVoice for local capture and correct code names
Use IraVoice when the speech-recognition stage should run on your own computer. On IraVoice Local, recognition runs on-device with NVIDIA Parakeet and no audio or text leaves the computer.
Click into Cursor’s agent input or terminal, hold Right Command on a Mac or Right Ctrl on Windows, and speak. IraVoice finds the Git project Cursor has open and spells its code names, so “use effect” becomes useEffect. It reads only names, never source text, and types the prompt without pressing Enter by default.