Online sound structure analysis based on generative model of acoustic feature sequences

Abstract

We propose a method for the online sound structure analysis based on a Bayesian generative model of acoustic feature sequences, with which the hierarchical generative process of the sound clip, acoustic topic, acoustic word, and acoustic feature is assumed. In this model, it is assumed that sound clips are organized based on the combination of latent acoustic topics, and each acoustic topic is represented by a Gaussian mixture model (GMM) over an acoustic feature space, where the components of the GMM correspond to acoustic words. Since the conventional batch algorithm for learning this model requires a huge amount of calculation, it is difficult to analyze the massive amount of sound data. Moreover, the batch algorithm does not allow us to analyze the sequentially obtained data. Our variational Bayes-based online algorithm for this generative model can analyze the structure of sounds sound clip by sound clip. The experimental results show that the proposed online algorithm can reduce the calculation cost by about 90% and estimate the posterior distributions as efficiently as the conventional batch algorithm.