How interview AI tells your voice from the interviewer's

Three ways to know who is speaking

1. Separate streams: microphone for you, system audio for the call — the most reliable method.
2. Diarization: a model clusters voices by pitch and timbre when everything arrives on one stream.
3. Meeting-app metadata: some platforms label the active speaker, but desktop tools rarely see it.

Why separate streams beat clever models

When your voice comes only through the microphone and the interviewer's only through playback, attribution is free: there is nothing to infer. That clean split is what lets an interview tool ignore your own clarifying questions and react only to the interviewer. It is also why the same tool can behave brilliantly on headphones and erratically on speakers — with speakers, the interviewer's voice leaks back into your microphone and appears on both streams.

What diarization does, and its limits

Diarization answers “who spoke when” from a single mixed recording. It works by turning short slices of audio into voice fingerprints and grouping similar ones. It is good enough for meeting notes written afterwards, where a mislabelled sentence is a minor annoyance. In real time it has less audio to work with and must decide immediately, so it mislabels more — particularly with two similar voices, short interjections such as “right” or “mm-hm”, and people talking over each other.

Setups that cause confusion

Panels are less of a problem than people expect. All interviewers arrive on the playback stream, and for answering purposes it rarely matters which one asked. What matters is that you are kept separate. For the specific case of wireless headsets, see Bluetooth headsets and interview audio, and for speakers versus headphones in general, headphones vs speakers.

Why your own voice should be heard too

Capturing your microphone is not only for attribution. In mock interviews it is how your answer gets transcribed and scored. In live use it gives the tool context — if you have already covered a point, the next suggestion can build on it rather than repeat it. Tools that ignore your audio entirely tend to suggest things you said thirty seconds earlier.

A five-minute audio check

1. Plug in wired headphones or a USB headset and set it as both input and output in the meeting app.
2. Set the same device as the system default, so the meeting app and the interview tool agree.
3. Join a test call and have someone speak; confirm their words appear as captions labelled as the interviewer.
4. Speak yourself; confirm your words do not trigger an answer.
5. Turn the call volume down to a comfortable level — loopback capture does not need it loud.

If your own speech is being treated as the interviewer's, audio is leaking from your output into your microphone. Headphones nearly always fix it. If the interviewer is not captioned at all, the tool is capturing the wrong output device — usually because the meeting app is playing through a headset while the system default is still the laptop speakers. Align the two and test again.

Privacy implications

Two streams means two people's voices are processed. Check how a tool handles that: whether audio is stored, where transcripts go, and whether anything is retained after the session. Recording or transcribing another person can also carry consent obligations depending on where you are. See AI interview assistant privacy.

Where CoPilot Interview fits

CoPilot Interview, a free AI interview assistant for Windows and macOS, captures your microphone and the call's system audio as separate sources on your machine, which is how it reacts to the interviewer rather than to you. Audio is streamed in short encrypted segments for transcription and is not stored on our servers.

Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly. Everything else on this page applies whichever tool you use, or none.

Frequently asked questions

How does interview AI know which voice is mine?

Mostly by keeping your microphone and the computer's playback audio separate. Your voice arrives on the microphone, the interviewer's on system audio, so the software does not have to guess.

What is speaker diarization?

Diarization is the process of working out who spoke when from a single mixed recording, by grouping slices of audio with similar voice characteristics. It is less reliable in real time than keeping speakers on separate streams.

Why does the AI respond to my own questions?

Usually because the interviewer's voice and yours are mixing — typically from laptop speakers echoing into the microphone. Wearing headphones restores the clean split between the two streams.

Does speaker separation work in panel interviews?

Yes. All panelists arrive on the call's audio stream, and for answering it rarely matters which one asked. What matters is that your own voice is kept apart from theirs.