Why architecture matters here

Diarization fails on overlap + unknown speaker count. Architecture matters because embed + cluster + fusion compose.

Advertisement

The architecture: every piece explained

The top strip is preprocess. Audio input. VAD. Segments. Speaker embed.

The middle row is grouping. Cluster. Change point. Overlap detection. Labels.

The lower rows are ops. ASR fusion. Metrics. Ops — segment size + threshold + rediarize.

Speaker diarization — embed + cluster + change point + labelswho spoke whenAudio inputmulti-speakerVADvoice activitySegments0.5-2sSpeaker embedx-vector / ECAPAClusterspectral / AHCChange pointspeaker turnsOverlap detectioncross-talkLabelsSPK-01, SPK-02ASR fusiontext with speakerMetricsDER + JEROps — segment size + threshold + rediarizeclusterdetecthandlelabelfusewatchwatchoperateoperate
Speaker diarization: VAD, embed, cluster, label.
Advertisement

End-to-end flow

End-to-end: audio in. VAD segments speech. 1s windows -> ECAPA embeddings. Spectral clustering with unknown-k. Change-point smooths labels. ASR aligned per segment. Output SPK-01 utterance start-end + text.