1. Use built-in Voice for a live Codex conversation
Open the ChatGPT desktop app on macOS or Windows, select Codex, then select the Voice control. Allow microphone access when prompted and describe the task or question.
OpenAI documents Voice in Codex as a live interface that can use the tools and permissions available to Codex. That makes it the natural choice for back-and-forth coordination.
2. Treat live Voice and dictation as different interfaces
Live Voice is conversational: you can speak, interrupt, ask questions, and coordinate an active task. Dictation is text-first: speech becomes editable text before you decide what to send.
Choose based on the control point you want. A live conversation is fast; a visible brief is easier to inspect, archive, compare, or reuse across agents.
3. Check the current Voice data controls
OpenAI’s Voice documentation describes retention and model-improvement controls for audio, video, and transcripts. These rules can depend on the voice mode, plan, and workspace, so use the current OpenAI help article rather than a third-party summary.
IraVoice keeps its speech recognition and supported formatting on the Mac, with the initial model download and paid license verification disclosed separately.
4. Use Voice-to-Spec for a reviewable handoff
Hold Globe in IraVoice and describe the goal, constraints, and acceptance criteria. Focused-app delivery inserts the resulting Markdown for review.
If you explicitly select Codex as the Prompt destination, IraVoice can open a visible interactive Codex session in the detected repository and submit a guard-accepted specification. It does not use unattended or permission-bypass flags, and it fails closed to the clipboard when the CLI, repository, focus, trust, sign-in, or guard state is not safe. Codex still owns execution and permissions.