How Does an AI Notetaker Know Who Said What?

Two jobs hiding behind one label

A name on a transcript line is the end of two separate jobs. First the software splits the audio into distinct voices, which is why an unnamed transcript shows Speaker 1 and Speaker 2. Then it has to decide which real person each voice belongs to. The first job is about sound. The second is a lookup, and the lookup is where tools really differ, because the name has to come from somewhere.

There are three places it can come from. A bot in the call can read the participant list the meeting platform already holds. A tool can learn what a person sounds like and recognise them the next time. Or software on your own Mac can watch the meeting window, where Zoom or Meet highlights whoever is talking. All three produce a tidy transcript on a good day. They go wrong in different ways, and only one of them leaves something behind about the people you met.

Borrowing names from the meeting platform

Fireflies is clear that its names come from the call, not from the audio. Its FAQ says uploaded files get generic labels like Speaker 1 and Speaker 2, and that speaker name detection only works when Fireflies joins a video meeting or when you use its Chrome extension. Otter describes its Notetaker the same way: on Zoom it tags speakers in real time based on the participant names in the meeting.

This route is only as good as the names people join with. Otter's own best practices page gives the example of one person showing up as John (ACME Solutions) in Teams, Johnny A. Smith on the calendar and Johnathan Smith in Otter, and asks you to make all three match. A meeting room that joins as a single tile is one name for everyone sitting in it. A dial in caller is a phone number. None of this is a defect in any one product. It is the ceiling of a method that can only repeat what the platform was told.

Learning voices, and what that leaves behind

The second route is the one that gets better over time, and Otter documents it in detail. Its Speaker Identification Overview says that tagging a speaker teaches Otter to recognise that speaker in future conversations, and its tagging page says each tag builds the voice identification for that person. The best practices page calls this speaker enrollment and says it builds a voice profile. It even suggests importing old recordings of the people you talk to most and labelling them in advance. Inside a Workspace, each member's speaker profile is shared, so a teammate is tagged automatically in a colleague's meeting.

That convenience has a cost the feature list does not mention. The people whose voices get profiled are mostly not the account holder. They are the client, the candidate, the investor and the vendor on the other end, none of whom created an account or agreed to anything. And removing one is harder than adding one. The Overview's FAQ answers the question of whether you can delete a speaker with a plain no, not at this time. A newer Speaker Management page, still rolling out, lets you merge, rename and hide speakers. Hiding takes a speaker off the tagging list, and the page says it cannot be undone. It does not say the voice data is deleted.

Why lawyers have started calling it a voiceprint

Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier. Before a company collects one, the statute requires it to tell the person in writing that it is collecting it, for what purpose and for how long, and to get a written release. It also requires a published schedule for destroying the data.

On 13 August 2026, Judge Eumi K. Lee of the Northern District of California refused to dismiss the BIPA claims in the Otter privacy class action. The plaintiffs allege that Otter stores each voiceprint so it can identify the same people in future meetings, and point to the speaker identification profile the tagging process creates. The court held that, taken as true, those allegations plausibly describe voiceprints used to identify individuals. That is a ruling on the pleadings. Otter has not been found liable for anything, and the facts are still to be tested.

It is not a one company issue either. Fisher Phillips reports a separate proposed class action, filed by Katelin Cruz in Illinois federal court on 18 December 2025, against an AI notetaker vendor on the same theory: a tool joined a meeting she never signed up for, identified speakers, and allegedly created and stored voiceprints without the notices BIPA requires. The firm's advice to employers is that tools which identify speakers or tie transcripts to individuals sit in biometric territory, and that the statute can reach a company based outside Illinois if a participant is physically in the state. If you sell to anyone in Chicago, the speaker identification method of your notetaker is now a procurement question.

How Driffle names speakers without a bot or a voice profile

Driffle takes the third route. There is no bot, so there is no participant list handed over by the platform. Instead, while you record, Driffle reads which person the meeting app marks as speaking, on Zoom, Google Meet, Microsoft Teams, FaceTime, Slack huddles, WhatsApp, Webex and Discord, including Teams and Zoom in a browser, and it uses your calendar invite's guest list. Each voice in the recording is then matched against the moments when exactly one person was highlighted.

The matching is built to refuse rather than guess. A voice gets a name only when most of its speech, 60 percent or more, lines up with one person's highlight and clearly leads the runner up. A moment when two tiles lit at once counts for nobody. A candidate name has to hold for seven seconds before it sticks, and once a voice is named it stays named for the rest of the call, because a name that changes halfway through is the error people remember. When Driffle is not sure, the line stays Speaker. It is never credited to somebody else.

Your own voice gets its own rule. If one voice is on your microphone, it is you. If a second voice shows up there, as it does when two people share a laptop in a room, Driffle stops assuming and offers a Who's this? prompt, and one tap names that voice for the rest of the recording.

What it does not keep is the point. Voices are grouped inside a single recording, and nothing about how anyone sounds is carried into the next meeting. There is no speaker profile to share across a team and none to delete later, and the audio itself is transcribed and discarded. The names in your notes come from the meeting window and the invite, not from a model of anyone's voice.

The limits are real. Reading the meeting window needs macOS Accessibility permission; it is optional, and without it you still get a full transcript but not the names. While capture is paused or off, names are not read from the window either. Being strict means you will see more lines left as Speaker than you would in a tool that takes its best guess, and you will fix some by hand. Driffle runs on Apple Silicon Macs as a direct download, with Intel Mac, Windows and Linux on a request form, and transcription and AI run in the cloud, so it is not an offline tool.

Four questions to ask before you choose

Put these to any vendor in writing, and keep the answers with your procurement notes.

• Where does each name on the transcript come from: the meeting's participant list, a stored voice profile, or something else?
• If the tool keeps voice profiles, who can see them, are they shared across the workspace, and can one named person's profile be deleted on request, by whom, and how fast?
• What does the transcript show when the tool is unsure: a generic label, or its best guess?
• If anyone on your calls might be in Illinois, where is the written biometric notice and the published retention and destruction schedule?

One test call worth more than the demo

Book a call with three people. Have two of them join from one laptop in the same room, and have the third join with a display name that is a company name rather than a person's. Agree one action item out loud from each person. Then read the transcript and the summary. A tool that labels the shared laptop as one person and hands that person both action items has shown you its method. A tool that leaves a line unnamed has told you the truth.

If you want names on your notes without a bot in the call and without a stored profile of anyone's voice, Driffle is a Mac download at driffle.ai/download.