An analytic study on clustering driven self-supervised speaker verification

Abderrahim Fathan, Jahangir Alam

Abstract: Highlights•We performed a fine-grained analysis of the quality and limitations of various clustering algorithms and configurations.•We find high correlation between clustering performance and the downstream self-supervised speaker verification performance.•Metrics like Completeness, Silhouette, and Davies–Bouldin score have the highest correlation with the downstream performance.•We suggest employing unsupervised clustering metrics as an alternative way to assess the generalizability of pseudo-labels.•Mixup is effective against label noise, and tends to be more helpful when the clusters are not compact or not well distanced.

External IDs:dblp:journals/prl/FathanA24