Effect of single-speaker x-vector training on overlapped-speech diarization

Determine whether training speaker-embedding extractors such as x-vectors exclusively on single-speaker utterances degrades embedding quality in overlapping regions and increases speaker confusion in overlap-aware diarization, and quantify whether training on overlapping or noise-augmented data mitigates this effect.

Background

In overlap-aware spectral clustering on AMI, missed speech decreased but speaker confusion increased. The authors hypothesize this is due to a mismatch between x-vectors trained on single-speaker utterances and overlapped test regions. They show partial improvement with noise-augmented x-vectors and visualize overlap embeddings with t-SNE, suggesting noisy overlap representations may drive confusion.

Validating or refuting this conjecture would clarify whether embedding training conditions materially affect overlap diarization errors and guide training data design for robust x-vectors.