Glossary
Speaker diarization
Splitting an audio recording into segments and labeling which speaker was talking in each, without necessarily identifying who they are.
Also called: speaker segmentation
Speaker diarization answers "who spoke when" in an audio recording: it segments the audio and labels each segment as speaker 1, speaker 2, and so on, without necessarily determining the speakers' actual identities. It is a distinct task from transcription — a diarization system can correctly separate two speakers' turns without understanding a word either of them said.
Diarization systems typically extract a voice-characteristic embedding for short audio windows, cluster windows with similar embeddings as belonging to the same speaker, and then merge adjacent windows into speaker turns. It is commonly run alongside, or piped into, automatic speech recognition, so a transcript can be attributed to individual speakers rather than presented as one undifferentiated block of text; the combination is what produces a labeled call or meeting transcript.
Speaker diarization matters for conversation intelligence platforms that analyze sales or support calls, for meeting transcription, and for any speech analytics pipeline where attribution to a specific speaker changes the analysis, such as measuring talk-time ratio. The main pitfalls are overlapping speech, where two people talk at once, and diarization error accumulating on top of transcription error, so a downstream analysis inherits mistakes from two separate models rather than one.
Last reviewed September 22, 2026