Speaker diarization v2 architecture — segmentation, embedding, clustering, streaming

Why architecture matters here

Diarization fails on overlaps, low-SNR audio, and streaming latency. Architecture matters because segmentation + embedding + clustering + overlap handling all impact DER.

The architecture: every piece explained

The top strip is the pipeline. Audio. VAD / segmentation. Speaker embedding x-vectors. Clustering.

The middle row is variants. Speaker turn detection. Streaming online. Overlap handling. Speaker enrollment.

The lower rows are ops. Confidence. Metrics DER + JER. Ops — languages + privacy + eval.

End-to-end flow

End-to-end: 3-speaker meeting. VAD segments. Embeddings extracted. Clustering identifies 3 speakers; enrollment matches known. Streaming updates as new audio arrives. Overlap detected in region; multi-label. Output: (speaker, start, end) tuples.