A new paper wants to show us what a speaker recognition neural network is actually doing inside.
Researchers applied two hierarchical clustering algorithms — Single-Linkage Clustering and HDBSCAN — to the internal representations of a speaker recognition network. Where earlier work used flat clustering methods like K-means to find simple groupings, this approach looks for nested, tree-like structures in how the network organizes what it hears. The team also built a new tool called Hierarchical Cluster-Class Matching (HCCM) to link those discovered clusters to human-legible categories: male, female, Irish, British, and combinations thereof. A scoring metric called the Liebig score quantifies how well each cluster matches its assigned category and flags what limits the match.
Speaker recognition is already deployed in phone banking, courtroom forensics, and law enforcement — contexts where an unexplained output has real consequences. Giving auditors a map of how a model internally groups voices by gender and nationality is a step toward accountability, even if the paper stops well short of prescribing fixes for what those maps reveal.
The irony worth noting: a system that can implicitly sort voices by nationality is also a system with the raw ingredients for demographic bias — and making the representations legible does not, by itself, make the model fair.