A density peak clustering approach to unsupervised acoustic subword units discovery

This paper studies unsupervised acoustic units discovery from unlabelled speech data. This task is usually approached by two steps, i.e., partitioning speech utterances into segments and clustering these segments into subword categories. In previous approaches, the clustering step usually assumes the number of subword units are known beforehand, which is unreasonable for zero-resource languages. Moreover, the previously-used clustering methods are not able to detect non-spherical clusters that are often present in real-world speech data. We address the two problems by a brand new clustering method, called density peak clustering (DPC), which is motivated by the observation that cluster centers are characterized by a higher density than their neighbors and by a relatively large distance from other points of a higher density in the space. Experiments on unsupervised acoustic units discovery demonstrate that our DPC approach can easily discover the number of subword units and it outperforms the recently proposed normalized cuts (NC) clustering approaches [1].