1. Enable Gemini CLI’s native Voice Mode first
Use a current Gemini CLI release. Google’s release notes say real-time Voice Mode arrived in v0.41.0, and the current settings reference still marks it experimental and disabled by default.
Open /settings, find Voice Mode under Experimental, and enable it. Then run /voice in the interactive CLI. Native voice is the shortest path when the Gemini prompt itself is the destination; no separate Mac dictation app is required for that job.
- Update Gemini CLI before looking for the setting.
- Open /settings and enable experimental Voice Mode.
- Run /voice in an interactive session.
2. Choose the transcription backend before speaking
The current default backend is gemini-live. Google’s settings reference explicitly says recordings using that backend are sent to Google Cloud for transcription.
Choose whisper when local transcription is the requirement. Keep the boundaries separate: local Whisper can keep microphone audio in the local transcription stage, but the resulting prompt still enters Gemini CLI and is processed under Gemini’s normal account, model, tool, and data controls.
3. Use push-to-talk or toggle recording deliberately
The default activation mode is push-to-talk. Hold Space while Voice Mode is active, speak, and release when the prompt is complete. Google exposes the app.voiceModePTT action in its keyboard-shortcut reference.
The alternate activation setting is toggle. Use it when holding Space is awkward for a longer requirement, then stop recording explicitly. In either mode, review the transcript before submitting agent instructions that contain commands, paths, permissions, numbers, or negations.
4. Match the local Whisper model to the job
Google currently exposes /voice model and a Whisper Model setting for the local backend. The documented choices range from smaller English-only models to quantized large-v3-turbo models.
A smaller model reduces the local footprint; a larger model can demand more storage and compute. Do not describe the backend as universally offline: model acquisition, Gemini authentication, and the later model request remain separate network and service boundaries.
5. Do not confuse transcription privacy with agent privacy
Voice transcription creates prompt text. Gemini CLI can then send that text, selected context, tool results, and later turns through the configured Gemini service. A local speech backend does not turn the whole coding agent into a local application.
Read the active workspace settings as well as the user settings because Google documents that workspace settings override user settings. That matters when a repository changes experimental.voice.backend or other Voice Mode behavior.
6. Add IraVoice only for the cross-app or brief workflow
Use Gemini CLI Voice Mode when you are already at its prompt and want either its cloud backend or its optional local Whisper path. Native input owns that destination and stays closest to the current CLI state.
Use IraVoice when the same held-key habit should work in supported email, notes, browser, and developer fields, or when a spoken requirement should become a reusable Markdown brief before you choose the destination. IraVoice has no direct Gemini CLI destination; review the focused-app or clipboard output and paste it into Gemini CLI yourself.