Accelerated Nonparametric Bayesian Double Articulation Analyzer for Unsupervised Word Discovery

This paper describes an accelerated nonparametric Bayesian double articulation analyzer (NPB-DAA) for enabling a developmental robot to acquire words and phonemes directly from speech signals without labeled data in more realistic scenario than conventional NPB-DAA. Word discovery and phoneme acquisition are known as important tasks in human child development. Human infants can discover words and phonemes from raw speech signals at eight months without any label data, unlike supervised learning-based speech recognition systems. NPB-DAA was proposed by Taniguchi et al. and shown to be able to perform simultaneous word and phoneme discovery without any label data. However, the computational cost of NPB-DAA was extremely large, and thus could not be applied to large-scale speech data. In this paper, we introduce lookup tables for conventional NPB-DAA to reduce the computational cost and developed an accelerated NPB-DAA. Using the lookup tables, values calculated in each subroutine are memorized and reused in the subsequent calculations. This acceleration does not harm the quality of word and phoneme discovery because the introduction of the lookup tables is theoretically supported. This paper also shows that our accelerated NPB-DAA significantly reduced the computational cost by 90% compared to conventional NPB-DAA.