Abstract

Word discovery from speech signals is a crucial task for a human infant to learn a language. Differently from conventional approach towards automatic speech recognition, infants cannot use labeled data, i.e., transcribed text. They have to discover words from speech signals and learn meanings of the words in an unsupervised manner. We have been developing machine learning methods that enable a robot to learn words automatically. In this talk, I am introducing two unsupervised machine learning methods. One is for simultaneous learning of lexicons and object categories using multimodal latent Dirichlet allocation (MLDA) and nested Pitman-Yor language model (NPYLM). The other is nonparametric Bayesian double articulation analyser (NPB-DAA) for learning phonemes and words directly from speech signals using hierarchical Dirichlet process hidden language model (HDP-HLM). The both methods are based on Bayesian nonparametrics. I am also introducing our research field called symbol emergence in robotics.