Double Articulation Analyzer with Prosody for Unsupervised Word Discovery

Abstract

Human infants discover words and phonemes using statistical information and prosody. For unsupervised word discovery, Taniguchi et al proposed the Nonparametric Bayesian Double Articulation Analyzer (NPB-DAA) which was able to segment speech data into word sequences. However, NPB-DAA uses only statistical information such as the mel-frequency cepstrum coefficients. In this paper, we extend NPB-DAA method using prosody, i.e., Prosodic DAA, for unsupervised word discovery. We use the second order differential of the fundamental frequency and the duration of silent as the prosody. We show in an experiment that Prosodic DAA outperforms NPB-DAA.