How do AI note takers tell speakers apart in a panel interview?

Most AI note takers do not recognize voices on a panel call; they borrow names from the meeting platform. When three or four interviewers each join Zoom, Microsoft Teams or Google Meet from their own device, tools that work through the platform, such as Otter's Notetaker or Fireflies' bot, label each speaker with the participant's name. Bot-free desktop tools hear one mixed stream of meeting audio, so they either read the meeting app's active-speaker indicator (Granola's speaker tags) or label only "me" and "them". In a room where the panel shares one microphone, tools fall back to "Speaker 1, Speaker 2" unless interviewers have enrolled voice profiles, as Teams Rooms and Zoom Rooms support, and the candidate, who is never enrolled, appears as an unnamed speaker. Facts below are as of November 2026.

We make Interview Signal, one of the tools below; everything about the other products comes from their own sites, linked and dated. For choosing a note taker for panels on other grounds (duplicate bots, independent scorecards, price), see best AI note taker for panel interviews; for the underlying technique, see what is speaker diarization.

How does an AI note taker know who is speaking?

There are three methods, and most tools combine two of them.

1. The meeting platform says who is talking. Each person who joins from their own device has their own connection, so Zoom, Teams or Meet knows whose audio is whose. A note taker that joins the meeting, or reads the meeting app, can attach that participant name to the words.
2. The audio arrives on separate channels. A desktop app can record your microphone and the computer's sound separately. That tells "you" from "everyone else" with certainty, but nothing more.
3. The software clusters voices. Diarization groups stretches of speech that sound like the same person into Speaker 1, Speaker 2 and so on. Names come only from manual tags or from voice profiles that people enrolled in advance.

Microsoft's Teams Rooms documentation (dated August 27, 2026) explains why method 1 breaks in a room: without speaker recognition, "the video, and audio feed for people in the room would be attributed to the space (for example, Conference Room 1), not the individuals speaking."

Which method does each tool use?

As published by each vendor, as of November 2026. "Not stated" means the vendor's pages we read do not say.

Sources: Zoom on audio transcription, per-participant audio files and smart name tags; Microsoft's Teams Rooms voice recognition; Otter's speaker identification overview and Disable Speaker Learning; Fireflies' joining FAQ and bot-free desktop guide; Granola's transcription and speaker tags pages; Interview Signal's own app.

What happens on a remote panel with three or four interviewers?

This is the easy case, as long as everyone joins from their own laptop. A platform-based tool gets a separate participant for each interviewer and for the candidate. Otter's help center says its Notetaker "will automatically tag speakers in real-time based on the participant names of Zoom meetings", and Fireflies says speaker names are detected when it "joins a video meeting". Granola's speaker tags use "the participant names and active-speaker information shown by your meeting app", read through the operating system's accessibility features.

Where remote panels still go wrong:

• Two interviewers on one laptop. Granola says that "if several people speak through the same shared device, Granola cannot distinguish the individuals using it", and the same is true of any tool that relies on the participant list.
• Interruptions and crosstalk. Granola warns that with overlapping speech it "may not be able to assign every part of the conversation to the correct person." Panels that interrupt the candidate get worse transcripts.
• A bot-free tool without platform cues. Fireflies' desktop mode shows "Speaker 1 / Speaker 2 instead of named participants." Interview Signal separates only your microphone from the meeting's sound, so in your copy every other interviewer's question is labelled as the candidate. Tracking of your own questions and the panel's split of questions still work, but read quotes with that in mind; our Interview Signal panel guide covers how each interviewer's copy works.
• Phone dial-ins. Zoom says the audio "for all users who join by phone will be merged into one audio file", so a panelist on a dial-in loses their own track.

What happens when the panel sits in one room?

One microphone hears everyone, so no platform name is available. Three things can still produce named speakers:

• Teams Rooms Intelligent Speaker. Microsoft says it can "distinguish between different speakers in the room" and "identify them", but only people "enrolled with their voice profile" who were invited to the meeting; people who aren't enrolled show as "Speaker". It needs a Teams Rooms Pro license for a room, or, for a laptop with a room speaker, a room host with Teams Premium or Microsoft's AI assistant license, and Microsoft suggests limiting the room to 10 in-person attendees for the best precision.
• Zoom Rooms smart name tags. Users "enroll and consent through their Zoom account" by recording a voice sample. Zoom says tags "perform best when applied to 1-16 in-room participants", and manual tags "are not saved when the meeting ends."
• Learned voices. Otter groups voices into Speaker 1, Speaker 2 and so on; each time you tag one, Otter "will be able to recognize that speaker in future conversations." Tagging can only start after the recording stops.

In every one of these, the candidate is the unnamed voice. That is fine as long as someone renames the label before quotes go into a scorecard. Interview Signal does not attempt it: in a room, its pre-interview screen says both voices reach the same microphone and labels will not be reliable. Our in-person interview note taker guide covers room setups in more detail.

Are voiceprints a privacy problem in interviews?

They can be. Microsoft says its speaker recognition analyzes "pitch, tone, and speaking style" to create "a voiceprint for each participant", stored in your organization's tenant. Zoom stores enrolled users' reference audio in the Zoom cloud. Otter's admin setting exists, in Otter's words, for "organizations with privacy, security, or compliance requirements related to voice data and speaker recognition", and turning it off has to be requested through an Otter account manager.

Illinois' Biometric Information Privacy Act lists "voiceprint" in its definition of a biometric identifier, and it requires an informed written release before collection. Enrolling your own interviewers is an employee-consent question; tagging a candidate in a tool that learns voices may create a voice profile of someone who never agreed to one. Our BIPA and hiring tools page explains the rules. This is not legal advice.

How do you check speaker labels before a panel debrief?

1. Run a test panel with colleagues, one of them playing the candidate, using the exact setup you will use: same rooms, same laptops, same tool.
2. Ask each interviewer to join from their own device for remote panels, with their own microphone on, even if two of them share an office.
3. Agree one capture tool per panel so labels are fixed once, not reconciled across several transcripts.
4. Rename generic labels before anyone copies a quote. A quote attributed to "Speaker 3" is not evidence about the candidate until someone checks who Speaker 3 is.
5. Score from what the candidate said. Before a quote supports a rating, confirm it came from the candidate, not from a panelist's question; the panel interview plan template assigns who covers which competency, which makes this check faster.

Which setup should a panel use?

• Remote panels, bot acceptable: a platform-based note taker such as Otter's Notetaker or Fireflies, with every interviewer on their own device.
• Remote panels, no bot: Granola with speaker tags on the Zoom or Teams desktop app or Meet in Chrome, or a tool where each interviewer keeps their own copy, such as Interview Signal.
• Onsite panels in a Microsoft or Zoom room: Teams Rooms or Zoom Rooms speaker recognition for enrolled interviewers, with the candidate's label renamed afterwards and a policy on voice enrollment.
• Onsite panels with no room system: accept Speaker 1, Speaker 2 labels and rename them, or have one interviewer take written notes; see AI note takers for group interviews.