Speaker diarization: telling voices apart, then naming them
Speaker diarization splits a recording by who is speaking, so each line of the transcript carries a label: Speaker 1, Speaker 2. It tells voices apart; it does not know whose they are. A name comes from a person who recognises the voice, or from a guess based on what was said. Dalm, a Mac app for calls and meetings in the room, tells the voices apart on your Mac, and no audio is sent anywhere; the transcript, each line with its time and voice number, is synced to our servers. Naming is optional and holds for one meeting: in the next, the same person is a new voice again. There is a Free plan.
Who spoke when, and who they are: two separate questions
Three things go by similar names, and they answer different questions.
- Speaker diarization — who spoke when. The recording is cut into stretches of speech, and stretches that sound like the same voice are grouped together. What comes out is labels, not names.
- Naming — who each voice is. A person says who it was, or a guess is offered and a person confirms it.
- Speaker identification — recognising a voice heard before. A voice is kept from an earlier recording and looked for in new ones.
Dalm does the first on your Mac and asks you for the second, with a guess from a language model that reads what was said. It does not do the third: a voice is told apart within one meeting and is not compared with the voices of other meetings. That is why a transcript can show who said what without knowing who anyone is.
Two streams on a call, one microphone in a room
On a call, Dalm records two streams: your microphone and the sound of the call. Your voice is already known from the first one — the naming step says so: “The first voice is already known — the microphone of this Mac.” — and in the transcript it shows as “You”. Only the other voices have to be told apart, and they all come in the call’s sound. In an interview, that split is what keeps the questions apart from the answers — transcribing an interview.
In a room there is one microphone and one sound, so every voice, yours included, is told apart from it, and the app asks which one is you: “That’s me”. A phone call on speaker goes the same way. Without headphones, the other side’s voice from your speakers is taken out of your microphone once the recording ends, and when two people talk at once, both are kept. How a call and a room get recorded in the first place is on the page about meeting recorders. A file brought in as a meeting is split the same way as a room, which is how a podcast episode with several voices comes back voice by voice.
How Dalm puts names to the voices
Until you answer, each voice that has no name yet shows either a guess or a number — “Speaker 1”, “Speaker 2”. The guess comes from the language model that writes the summary: it reads the transcript, which is synced to our servers, and suggests names from what was said, not from how anyone sounds.
The naming step is called “Who was speaking” and goes voice by voice: “Voice 2 of 6”. For each voice it shows the longest line that voice said and when. A guess may arrive marked “Suggested by Dalm — check before you agree.” In the transcript and the summary a guess looks like any other name until you confirm or change it. You can type a name and a role, pick someone “From the invited”, answer “That’s me”, “Skip”, or “Name later”. Naming is optional: the summary is written without waiting for it.
If the app has cut one person into two voices, right-click the name in the transcript and choose “Merge With”. The “Who was speaking” list also lets you rename the voices. All of this is on the Free plan.
A name is given once per meeting, not once for good
A name you give shows up on every line of that voice in the meeting’s transcript, and it stays with that meeting on every Mac you sign in to: it is text, and it syncs. What it does not do is follow the person. In the next meeting the same colleague is a new voice again, on this Mac or any other, because a voice is told apart inside one meeting and is not carried over to the next. The language model may suggest the name again from what is said; the voice itself is not recognised. A weekly call means naming the regulars each week — or leaving them as Speaker 1 and Speaker 2, since the summary is written either way.
What is kept about a voice, and where
The audio is deleted once the text exists, and nothing of the sound is kept for naming. The transcript is different: it is synced to our servers automatically. On a call your own voice is not asked about; in a room or from a file it is asked about like any other voice.
To tell voices apart, the app makes a voice print for each one — a set of numbers describing how that voice sounds. It stays on this Mac for as long as the meeting exists, naming the voice does not erase it, and it is never synced. What reaches our servers is text: the transcript, each line with its time and voice number, synced automatically, and the names you give the voices. When a summary, a document or an answer is written, a language model receives that text and those names; the audio is not part of what it gets. Every kind of data, row by row, is on how Dalm works.
How accurate speaker diarization is, and why we give no percentage
How well voices are told apart is usually measured as a diarization error rate: the share of speaking time that was missed, invented, or given to the wrong voice. We give no figure for Dalm: we have not published a measurement together with its method — how much speech, which recordings, who marked the reference — and a percentage without that says nothing about your meetings.
What you can see instead is where it went wrong, and put it right:
- one person cut into two voices — “Merge With” joins them;
- a wrong name — rename the voice in “Who was speaking”;
- a point in the summary whose speaker is in doubt carries the mark “Attribution uncertain”;
- a name the model suggested is its guess until you confirm it in the naming step.
Before a line goes into minutes with someone’s name on it, check who said it — how to write meeting minutes covers why that matters.
Free, Pro or Team: the first meeting with names in it
Telling voices apart and naming them come with the Free plan, and Free is a plan, not a trial: there is no end date. Dalm needs a Mac with Apple Silicon · macOS 15.0 or later and an account, and while the app is not public, the way in is the waitlist. What the paid plans add is on the pricing page. Who in a team sees a shared meeting is set out on the page for teams.
Questions
- What is speaker diarization?
- It is the step in transcription that splits a recording by who is speaking and gives each stretch a speaker label — Speaker 1, Speaker 2 — so the transcript reads as a conversation. It tells voices apart; it does not say whose they are. In Dalm the voices are told apart on your Mac, and the transcript with its voice numbers is synced to our servers.
- What is the difference between speaker diarization and speaker identification?
- Diarization answers who spoke when, and hands back labels. Identification answers whose voice it is, by comparing it with a voice heard before. Dalm does diarization on your Mac, and the transcript is synced to our servers; the names come from you, with suggestions from a language model that reads what was said. A voice is not recognised from one meeting to the next.
- Why does my transcript say Speaker 1 instead of names?
- Because telling voices apart and knowing whose they are are two different steps, and the first one only produces labels. In Dalm a voice you have not named shows the model's guess or a number. Open "Who was speaking" and give it a name. Naming is optional: the summary is written without it.
- Does Dalm recognise the same person in the next meeting?
- No. Voices are told apart within one meeting, and in the next one the same person is a new voice again, on this Mac or any other. The names you gave stay with the meetings you gave them in, and the language model may suggest a name again from what is said.
- Does Dalm send my voice to a server when it separates speakers?
- No audio is sent anywhere, and the voice prints used to tell voices apart stay on this Mac and are never synced. What goes to our servers is text: the transcript, each line with its time and voice number, and the names you give. A language model receives that text when it writes a summary, a document or an answer.
- How accurate is speaker diarization?
- It is usually measured as the share of speaking time that was missed, invented or given to the wrong voice, and it depends on the recording. Dalm gives no figure for it: we have not published a measurement with its method. When a voice comes out wrong, you can merge two voices into one or rename one, and a point in the summary whose speaker is in doubt is marked "Attribution uncertain".
- Is there a free app for Mac that labels speakers in meeting transcripts?
- Dalm has a Free plan, and telling voices apart and naming them are part of it. It is a plan, not a trial. You need a Mac with Apple Silicon · macOS 15.0 or later and an account, and while the app is not public, the way in is the waitlist. What the paid plans add is on the pricing page.