Give yourself half an hour the first time. Audio capture is the one thing in the product that needs setting up outside the browser. Walk through it before the day.
Three ways it hears the interviewer
Which one depends on whether you wear headphones and whether the call runs in a browser or a desktop app.
Through your speakers
Nothing to set up. A microphone listening to speakers captures both the interviewer and you, so silence cannot reliably identify whose turn it is. Automatic silence triggering is therefore unavailable on this route.
The default is a voice phrase: say “Let me think” after the interviewer finishes. The phrase is removed and only the preceding question is sent to the model. You can switch to manual in the live panel and use the shortcut or “Generate suggested answer” instead. Once generated, the suggested answer is locked: your reply is logged as “Me” and cannot replace it or start a second request. When you finish, choose “Continue to next question”.
With headphones the assistant cannot hear the other side; use one of the isolated audio routes below instead.
The call runs in a browser tab
Headphones are fine. When you start the live assistant, the browser asks which tab to share: choose the meeting tab and tick "Also share tab audio".
Headphones, with the call in a desktop app
When you start the live assistant, the browser asks what to share — choose "Entire screen" and tick "Also share system audio". On macOS this needs Chrome 141 or later. On older versions, install a virtual audio device such as BlackHole first, then select it as the input device.
Four things to confirm before you start
- The browser is Chrome or Edge. Nothing else can listen.
- Microphone access is granted. Without it the setup screen says so and offers a retry.
- The right capture route is selected, and audio sharing is ticked when you start.
- Your balance covers a session. Forty-five minutes runs about 450 credits.
A second screen
Prompts can be pushed to another window, away from the screen you are sharing. That window has to be another tab or window of the same browser: the two stay in sync inside the browser and nothing goes over the network. A phone is a separate device and will not receive the session.
Keyboard shortcuts
The setup screen lists the keys for four actions, all rebindable: generate a suggested answer for the current question, hide or show prompts, adjust overlay background opacity, and open second-screen prompts.
Triggering and pauses
- Speakers: voice phrase or manual only; silence auto-trigger is unavailable.
- Shared tab and system audio: these routes contain only the other side, so they offer auto or manual. Auto has quick, standard, and relaxed pause presets (1.0, 1.6 and 2.2 seconds) and can be changed during the interview.
- New transcription never interrupts a request already generating. A confirmed next question is queued.
- If a sentence was split too early, the live panel offers to merge it into the current question; it does not erase the existing answer and rerun by itself.
History and corrections
Every question, suggested answer, and follow-up is kept as an append-only list. Earlier turns expand from the prompt card and live panel, and the second screen has previous/next controls. If a question was misheard, choose “Edit question”, edit it, and select “Save and regenerate”. That explicit action is the only one allowed to stop an in-flight incorrect request; ordinary speech and new questions cannot.
Transcription runs on our servers
Signed in, transcription is done by our servers, which means interview audio passes through them. Confirm that this is acceptable where you are before you start. See Boundaries and data.