Non-Parametric Bayesian Subspace Models for Acoustic Unit Discovery

This work investigates subspace non-parametric models for the task of learning a set acoustic units from unlabeled speech recordings. We constrain base-measure Dirichlet-Process mixture with phonetic subspace---estimated other source languages---to build an \emph{educated prior}, thereby forcing learned to resemble phones known languages. Two types are proposed: (i) Subspace HMM (SHMM) which assumes that is same every language, (ii) Hierarchical-Subspace (H-SHMM) relaxes this assumption and allows have language-specific estimated on target data. These applied 3 languages: English, Yoruba Mboshi they compared various competitive discovery baselines. Experimental results show both outperform systems in terms clustering quality segmentation accuracy. Moreover, we observe H-SHMM provides superior SHMM supporting idea priors preferable language-agnostic unit discovery.