Answers are grounded in this paper and cite their sources.
Speaker diarization—a problem of segmenting and identifying speakers in audio recordings—is critical for applications such as automatic speech recognition, transcription, and analysis of multivoice recordings, including meetings and podcasts. Overlapping speech, which is widely represented in popular datasets, poses a significant challenge as it complicates accurate speaker segmentation. In this paper, this issue has been addressed by studying the linearity of biometric embeddings, a property that allows overlapping speech to be represented as a linear combination of individual speaker embeddings, which is key to robust diarization, especially in cascaded schemes with target speaker voice activity detection. A new method for retraining a speaker identification model aimed at improving the linearity of embeddings using a synthetic dataset constructed from a dataset of speaker voices and a modified loss function with an additional term responsible for linearity has been proposed. The integration of the proposed approach into a cascade-based diarization system provides support for both full context mode and streaming mode. Experiments on standard benchmarks demonstrate a reduction in the diarization error rate compared to state-of-the-art methods, indicating improved handling of overlapping speech. The proposed method addresses a gap in the field of embedding linearity optimization and has practical significance for real-world multivoice scenarios.
No takes yet. Share an insight, caveat, or question.
Popov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: