Why meeting transcripts get the speaker wrong, and how to get who said what right

A meeting transcript credits words to the wrong person when the information it uses to tell speakers apart runs out: one connection shared by several people, two voices at once, voices that sound alike, or a reply too short to place. This guide is for anyone who relies on a Google Meet or Microsoft Teams transcript, the platform's own or a notetaker's, to know who said what. It explains how speaker labels are produced, why each method fails where it does, the habits that prevent most mistakes, and what to check before minutes go to a client.

How does a transcript know who is speaking?

In one of two ways, and most errors come from the gap between them. Either the meeting platform says which participant's connection was sending sound, or software listens to a mixed recording and groups the voices it hears.

Method	How it works	Good at	Fails on
From the platform	Google Meet or Teams knows whose microphone is live and labels the line with that participant's name	Real names, remote participants on their own devices	Several people on one connection; overlapping speech
From the audio (diarisation)	Stretches of speech that sound alike are grouped into Speaker 1, Speaker 2	Recordings with no participant list	Similar voices, short replies, someone who changes microphone

Captions and the platforms' own transcripts start from the first method. A recording made on a phone in the middle of a table can only use the second, and then a person, or a model, has to decide which "Speaker 2" is which name. That last step is a guess, however well it is made.

Why does a shared meeting room break speaker labels?

Because the room is one participant. When four people sit around a room system or one laptop, the platform sees one connection with one name, and everything said in that room is credited to it: the room's name, or whoever owns the laptop. The transcript is not wrong about the connection. It is silent about the person.

There are three ways round it:

1. Join individually. Where people can, each joins from their own device with a headset.
2. Say your name. In a shared room, a person says their name before a point that matters: "Tom here, I'll take the migration."
3. Use the platform's room features. Microsoft documents a way for Teams' own transcript to tell apart and name people sitting in a Teams Room: each person sets up a voice profile in the Teams desktop app, and anyone without one appears as a numbered speaker labelled with the room. Availability depends on the room and the organisation's settings, so ask the client's IT team. The details are on Microsoft's support page.

Why do crosstalk and short replies end up with the wrong person?

Because there is no clean moment to attribute. When two people talk at once, the platform may see two live microphones and the recording has two voices in one stretch. Whatever the method, one line has to be split or given to one of them.

Short replies are the quieter version of the same problem. "Yes", "agreed" and "fine by me" are over before most methods can place them, and they usually land on whoever spoke just before or just after. That matters more than it sounds: agreement is how decisions are made in meetings, and a transcript that credits "agreed" to the wrong person has recorded a different decision.

Timing adds to it. Captions are produced live and can arrive a moment late. A tool that lines words up against a timeline of who was speaking can hand the first words of a reply to the previous speaker, especially when turns are quick.

Why do similar voices get merged, or one person split in two?

Because diarisation groups by sound, not by identity. Two colleagues with similar pitch and accent can become one "Speaker 1". One person who moves from a headset to a laptop speaker halfway through can become two, because they now sound like someone else. Echo from someone listening on speakers without headphones can put a second copy of a voice on another connection.

The platform method avoids most of this for remote participants, which is why it is the better source of names whenever it is available. It still fails when two people share a connection, and when someone joins twice, from a laptop and a phone, and appears under two names.

Why does AI sometimes name the person being spoken to?

Because the name in the sentence is the strongest clue in the text, and it usually belongs to the listener. "Thanks, Tom, I'll send it tonight" is said to Tom, by someone else. A model that guesses speakers from the words alone, with no timing to go on, tends to give that line to Tom, and the action item to the wrong person with it.

This is why speaker labels should come from evidence about who was talking, not from what was said. When the evidence is missing, an honest transcript leaves the name blank.

What can you do before and during the meeting?

Most speaker errors are prevented by habits that cost nothing:

• One person per connection wherever possible, with a headset.
• Real display names. Ask people to join under their name, not "iPhone" or a meeting room's asset tag.
• Names with commitments. "Priya will send the spec by Friday" survives any transcript; "I'll do it" depends on the label.
• A beat between turns. Waiting for someone to finish is better manners and a better transcript.
• Mute when not speaking, especially on laptop speakers, to keep echo off other connections.
• A read-back at the end. "So: Priya, the spec by Friday; Tom, the migration plan next week." It gives every notetaker, and every reader, one clean statement of who owns what.

What should you check before the minutes go out?

Not the whole transcript. Check the lines where a wrong name changes what the minutes say:

1. Every action item's owner. Would that person have said that?
2. Decisions and agreements. Who agreed, and was it the person with authority to?
3. Anything attributed to the client. A client misquoted in writing is the error most likely to be noticed.
4. Lines with no name, or a room's name, near any of the above.

If a line matters and you are not sure, listen to the recording at that moment. Names that are spelled wrong, rather than given to the wrong person, are a different problem, covered in why transcription gets names and jargon wrong.

How does TrackHeros decide who said what?

It takes names from the meeting platform and words from the recording, and it does not guess. The transcript is made in two passes: Google Meet or Microsoft Teams captions first, which carry the real speaker names, then an audio pass over the recording that produces the authoritative words. Each line of the audio pass is placed on the captions' timeline and takes the name of whoever the platform says was speaking at that moment. The same split is what keeps names right in a meeting that switches language mid-sentence.

Where that evidence runs out, in crosstalk, a long overlap or a stretch with no captions, the line is labelled "Unknown" rather than handed to the most likely person. Why a transcript says Unknown covers what to do with one. In the meeting minutes, an action item's owner is the name as heard; if that name did not match anyone in the meeting, the item carries a "Not an attendee" chip: "This name did not match anyone recorded in the meeting — verify before acting on it." Owners can be edited on the meeting page, and the recording plays inside the app for anything you need to check.