Bayesian non-parametric matrix factorization for discovering words in spoken utterances

In earlier work, we have formulated word discovery from speech as a latent component analysis problem. In more recent work, we proposed a Bayesian approach for estimating the model order, i.e. the vocabulary size, by evaluation of the marginal likelihood for different order values. That technique was expensive since the algorithm should be repeated for several order values to estimate the proper order. Here, we develop a Bayesian non-parametric approach to decompose the spoken utterances into word models. The number of latent components can be automatically discovered through taking a large number of latent components for the model and a weight vector representing the overall gain of each component. A sparse prior for the gain parameters lead to a limited number of activated model components. The word representations as well as the model order are then obtained from a single iterative variational Bayesian inference procedure, which constitutes a substantial advantage over the previous approach of trying multiple model orders. Experiments are performed on synthetic as well as real speech data.