Скачать или смотреть Fast and Accurate Speaker Diarization with Dedicated Overlap Handling

Speaker: Viktorya Poghosyan (CAST)
Topic: Fast and Accurate Speaker Diarization with Dedicated Overlap Handling
DataFest Yerevan 2025, https://datafest.am/

Abstract: Speaker diarization is the task of partitioning audio recordings into speaker-homogeneous segments, answering the essential question of "who spoke when?" This work focuses on developing an efficient speaker diarization system that maintains high accuracy while outperforming existing open-source solutions in terms of execution time. The system is specifically tailored to handle overlapping speech and dynamically changing numbers of speakers, without relying on prior knowledge of how many participants are present.
A new approach is proposed that uses iterative cosine similarity comparisons between segment-level speaker embeddings and a dynamically updated vocabulary of identifiers. To enhance accuracy in overlapping segments, a modified Transformer-based classifier is trained to identify active speakers. Extensive post-processing steps, including segment merging, re-labelling, and VAD-based corrections, further refine the diarization output.