Graph clustering beats K-means for recovering natural word distributions

Unsupervised term discovery — the task of segmenting unlabelled speech into word- or syllable-like units and clustering them into a lexicon — has long relied on centre-based clustering methods such as K-means. However, these methods carry an inductive bias toward spherical clusters, producing a more uniform distribution of type frequencies. True lexicons follow a Zipfian distribution, where a few words appear very frequently and many appear rarely. A new paper by Danel Slabbert, Simon Malan, and Herman Kamper (arXiv:2606.10781) revisits graph-based clustering as a bottom-up alternative that naturally preserves this property.

Using segment embeddings connected by pairwise similarity and partitioned via the Leiden algorithm, the authors demonstrate substantial improvements over K-means, Gaussian Mixture Models (GMM), and BIRCH across both word- and syllable-level lexicon discovery. Tested on three languages, graph clustering consistently produced lexicons whose frequency distributions closely matched true Zipfian curves. Agglomerative clustering with average linkage also performed well but was computationally slower and offered less control over the resulting distribution shape.

The findings call into question the longstanding dominance of centre-based clustering for this task. By showing that a graph-based approach yields more realistic lexicons without sacrificing scalability, the work points toward a simple yet effective tweak for researchers building unsupervised speech recognition pipelines. Future work could explore combining graph clustering with modern embedding techniques to further improve segmentation accuracy.