Two voices sounding alike is a matching problem, not a listening problem.

A transcript with the speakers mixed up still has every word in it — the recording worked, the transcription worked, and the meeting is fully there in text. What's wrong is narrower and easier to miss at first glance: a handful of lines are attributed to the wrong person, confidently, so a decision one speaker made reads as though the other one made it. That's a different failure than a transcript with a voice missing from it entirely, and it has a different, much smaller fix.

What's actually wrong when this happens

Coii AudioNotes tells speakers apart by comparing how each voice sounds against the others in the same recording — pitch, tone, cadence, the pattern that stays consistent for one person and differs from everyone else talking. That comparison, called speaker diarization, runs on voice prints built from the audio itself, not from any directory of who's expected to be on the call — the mechanics of how the app tells voices apart in the first place are the same ones doing the work here. Most of the time it lands cleanly. A mixed-up transcript means the comparison landed on the wrong side for a few specific segments — the words are transcribed correctly, but the label above them points at the wrong person.

Why it's a matching problem, not a broken recording

Nothing about the audio itself is damaged when this happens — both tracks captured fully, and every word made it into the transcript. What went wrong is the step that decides which of the two (or more) voices a given segment belongs to, and that step has to make a judgment call on every short piece of speech, including the quick ones. A missing voice is a recording failure: a track never captured anything, and that's a different problem with a different fix. A mixed-up transcript is a labeling mistake on a complete recording, which is why the fix is a correction, not a re-recording.

Why some conversations are more prone to it than others

A slow, two-person call with one speaker at a time is the easiest case: long stretches of one voice give the matching plenty to work with, and the pause between turns marks the handoff clearly. The cases where a line gets crossed tend to share one of a few traits:

• Fast back-and-forth. A quick exchange — "you take that" / "got it" — gives the matching only a short segment to judge each turn on, which is harder than a full sentence.
• Voices that sound close together. Two people with similar pitch and pace, especially over a laptop's built-in microphone, give the comparison less to distinguish on than two very different-sounding voices would.
• Several people sharing one room mic. When three or four people are picked up by the same microphone rather than each having their own track, the app is separating that many voices from one mixed signal, which is a harder job than separating two tracks that were already kept apart.
• Overlapping speech. A moment where two people start talking at nearly the same time produces a segment that's genuinely a blend of two voices, and whichever one gets picked for the label, the other is briefly wrong.

None of these mean the recording did anything unusual — they're just the conditions under which telling two specific voices apart gets harder, the same way two people are easier to recognize by voice on a quiet phone line than across a noisy room.

The fix: correct the specific lines, not the whole transcript

Fixing a mixed-up attribution uses the same mechanism as correcting any other speaker label: click the label on the line that's wrong and set it to the right person. The difference from a generic placeholder correction is scope — you're not renaming "Speaker 2" everywhere it appears, you're moving one or two specific lines from one already-named speaker to the other, because the rest of the transcript already has both people labeled correctly. Find the crossed lines — usually a fast exchange where both lines land under the same name for a beat, or a single line reads as though the wrong person said it — and correct just those, leaving the rest of the transcript untouched. The words themselves never need editing; only the name above them does.

Why correcting it also helps the next meeting

Coii AudioNotes checks new recordings against the voices it's already matched, so a correction made now gives it a cleaner example of each person's voice to compare future segments against — the two speakers who got crossed once are less likely to get crossed again once the correction has sharpened the distinction between them. It isn't a guarantee against every fast exchange in every future call, but it's a real improvement, not just a fix for the one transcript in front of you.

What this does to the minutes

Because the generated minutes are built from the transcript, a wrong attribution can turn into a wrong action item — "Priya to take the rollback plan" when it was actually Marcus who agreed to it. That's the practical reason to fix crossed lines rather than shrug them off as a minor labeling quirk: the mistake doesn't stay contained to the transcript, it carries into the summary that people who weren't on the call are the ones most likely to act on without checking the source.

When it's worth checking the transcript closely

Not every transcript needs a line-by-line audit. A recurring one-on-one between two people whose voices are already well established is unlikely to need any correction at all. The transcripts worth a closer read are the ones with a fast, decision-heavy exchange between two similar-sounding voices — the exact moment where getting the attribution right matters most is also the moment it's hardest to get right automatically. Skim for the handful of places where a quick back-and-forth lands right before a decision or a commitment, and check those specifically rather than reading the whole thing start to finish.

Several people in one room versus a room and a call

The two-track setup — your own microphone and the call's system audio, kept separate — already does half the work of telling people apart before any voice matching starts, because it never has to separate you from the person on the other end of the call. The harder case is the room side of that split: if two or three colleagues are gathered around one laptop for a call, all of them are on the same microphone track, and separating three or four voices from one mixed signal is a meaningfully harder job than separating one voice from another that was never mixed together in the first place. A recruiter running back-to-back screening calls with a single candidate on the other end rarely sees this; a panel interview with three interviewers sharing one room and one mic is exactly where it shows up.

Telling a mixed-up speaker apart from a misheard word

A speaker attributed to the wrong person and a word transcribed incorrectly are separate mistakes, and it's worth not conflating them when reading back a transcript that has something wrong with it. A misheard word is the transcription getting the sound of the speech wrong — "Dana" heard as "Donna," a number misheard by one digit — and it sits inside a line that's otherwise attributed correctly. A mixed-up speaker is the opposite: the words are exactly right, transcribed cleanly, and what's wrong is which name is written above them. The first is a transcription accuracy question; the second is a voice-matching question, and the fix for one doesn't touch the other.

What a two-person call with very different voices looks like by comparison

It's worth knowing what the easy case looks like, if only to recognize when a transcript is actually at low risk of this. Two people with clearly different voices — different pitch, different pace, different accents — give the matching a wide margin to work with on every single segment, including short ones. A one-on-one between a manager and a direct report, recorded over a decent microphone with normal turn-taking, is about as unlikely to produce a crossed line as this kind of matching gets. The traits that raise the odds — speed, similarity, several voices on one track, overlap — tend to compound with each other rather than acting alone, so a fast exchange between two similar-sounding people in a crowded room is the case actually worth double-checking, not the ordinary two-person call.

What actually reduces how often this happens

There's no setting that eliminates voice matching mistakes entirely — any system comparing voices against each other is going to have edge cases where two of them sound close enough to cross. What reduces it in practice is mostly about the room: fewer people sharing one microphone, and a microphone close enough to pick up a clear, distinct signal from each speaker rather than a faint, reverberant one from across a table. None of that is a claim that the transcript will be perfect — it's a claim that a correction, once made, tends to be the last one that pair of voices needs.

If the same two people keep getting crossed

Most crossed lines are a one-off — a fast exchange in one particular meeting, corrected once and rarely repeated with that pair. If the same two people keep getting mixed up across several meetings even after corrections, that's usually a sign their voices genuinely sit close together on the traits the matching relies on, rather than any single meeting being unusual. Correcting each occurrence as it comes up still helps — the app has more captured audio to distinguish the two voices with every correction — but two people who sound unusually alike may need a correction more often than an average pair, the same way two people who look alike get mistaken for each other more than two people who don't, no matter how many times someone's told them apart before. Recording with the microphone a little closer to whoever's speaking, rather than one mic equidistant from everyone in the room, gives each voice a cleaner signal and narrows that gap somewhat.

What it costs

Coii AudioNotes is $19 once, for three of your own Macs, with a 30-day trial that needs no card and no account to start. It runs on macOS 13 Ventura or later, and every correction you make to a transcript — a wrong name, a crossed line, a placeholder that needs a real name — stays entirely on the Mac that made the recording, the same as the recording itself.