Speaker diarization is the "rows" to ASR's "cells": ASR tells you WHAT was said, diarization tells you WHO said each segment. Together they produce the interview-transcript format that every meeting-notes and call-centre product relies on.
Classical diarization used voice-activity detection + speaker embeddings (x-vectors) + clustering. Modern end-to-end neural diarization (EEND) predicts speaker activity jointly with segment boundaries, and handles overlapping speech better than the clustering approach. Production systems (Pyannote, NVIDIA NeMo, AssemblyAI, Deepgram) combine both, with Whisper-class ASR running in parallel and the outputs time-aligned.
The hard cases are **overlap** (two people talking simultaneously), **acoustic mismatch** (one speaker on a headset, another on speakerphone), and **long recordings with intermittent speakers** (one-hour meeting where someone speaks twice for 30 seconds). Error rates on clean two-speaker audio are under 5%; four-plus-speaker meeting audio still hits 10–20% diarization error rates on public benchmarks.
Where AIMenta applies this
Service lines where this concept becomes a deliverable for clients.
Beyond this term
Where this concept ships in practice.
Encyclopedia entries name the moving parts. The links below show where AIMenta turns these concepts into engagements — across service pillars, industry verticals, and Asian markets.
Other service pillars
By industry