What is speaker diarization?

What diarization adds to a transcript

Speech recognition turns audio into words. Diarization adds structure: it detects when the speaker changes and groups words spoken by the same voice under one label.

Google's Cloud Speech-to-Text describes it this way. It tags each word with a number for the speaker, and words from the same voice share the same number. Amazon Transcribe calls the feature speaker partitioning and labels voices spk_0 to spk_29.

A short example

Without diarization, a transcript reads as one block: "Can we ship Friday? Only if QA signs off tomorrow. I'll ask them today."

With diarization, the same audio becomes separate turns:

• Speaker 1: Can we ship Friday?
• Speaker 2: Only if QA signs off tomorrow.
• Speaker 1: I'll ask them today.

Why the labels aren't names

Diarization only compares voices with each other. Microsoft's Speech service, for example, returns generic IDs such as Guest-1 and marks some speech as Unknown. Turning Speaker 2 into "Priya" takes extra information.

That information can come from a meeting roster, a calendar invite or a person who corrects the labels afterwards. Video calls make this easier because the platform knows who is connected. A single microphone in a room has no roster.

What affects the labels

Diarization is an estimate, so check the labels before you rely on them. A few settings and conditions make the biggest difference:

• The expected number of speakers. Google asks for a minimum and maximum count, and Amazon takes a maximum. Set it too low and there are fewer labels than voices.
• Speech the service can't place. Microsoft's results can mark a turn as Unknown instead of guessing.
• Where the microphone sits in a room. Put one device where it can hear everyone, so quieter voices aren't lost.

Why it matters for meeting notes

Who said something often changes what it means. "I'll ask them today" is a commitment only if you know who said it.

Before you share a transcript or assign action items, check the labels around decisions and commitments. Fixing one wrong label can change who owns a task.

Speaker labels in Bridge

Bridge transcripts are speaker-labeled. In hosted and captured calls, names improve when Bridge knows the participants from calendar guests, workspace members and people in Memory.

For in-person recordings, voice AI labels speakers Person 1, Person 2 and so on. You can update the names on the transcript after the meeting.