[Bug] Live Voice transcribes same-machine playback audio as user speech

What version of the Codex App are you using (From “About Codex” dialog)?

26.727.51351 (build 6119, bundle identifier com.openai.codex)

Bundled runtime: codex-cli 0.146.0-alpha.9.2

What subscription do you have?

Unknown (not verified).

What platform is your computer?

macOS 26.5.1 (build 25F80), arm64.

What issue are you seeing?

During continuous Live Voice in the Codex desktop app, spoken media playing on the same Mac can be transcribed and handled as if it were fresh user microphone input instead of being filtered or cancelled.

In the observed session, the user told Codex to stay quiet and then watched a YouTube video without addressing Codex. Dialogue from the video appeared in the realtime task transcript as user input. A short phrase from the video (“every Lunchables flavor”) appeared in the transcript, and the captured passage was duplicated. Codex then began responding to the playback-derived input.

This can create unintended turns and allow unrelated machine playback to steer or add noise to an active Codex task.

What steps can reproduce the bug?

1. Open the Codex desktop app and start continuous Live Voice in a task.
2. Do not address Codex.
3. Play a YouTube video containing speech on the same Mac.
4. Observe the realtime transcript/task input.

Observed result: speech from the video is classified as user input and can trigger a Codex response.

This is a confirmed observation from one active session, not yet a controlled test across different microphones or output routes.

What is the expected behavior?

Live Voice should suppress or cancel audio originating from the same machine’s playback so that it is not promoted to user speech or used to steer the task. If reliable cancellation is not possible for a particular audio route, the UI should make that limitation clear and avoid automatically treating likely playback/echo as a new user turn.

Additional information

• Product area: Codex desktop app, continuous Live Voice (not one-shot composer dictation).
• The task transcript directly showed the YouTube dialogue classified as user input.
• Exact microphone, playback output route (speakers/headphones), volume, subscription, and voice model are unknown/not verified.
• Frequency across separate sessions is unknown.
• No matching open or closed issue was found after searching for system/machine playback, YouTube audio, speaker audio, loopback, echo cancellation, and noise suppression.
• Related but distinct: #12902 concerns an internal transcript re-ingestion/usage loop, not physical or system playback entering voice input.