Gibbs sampling based Multi-scale Mixture Model for speaker clustering

Abstract

The aim of this work is to apply a sampling approach to speech modeling, and propose a Gibbs sampling based Multi-scale Mix ture Model (M 3 ). The proposed approach focuses on the multi-scale property of speech dynamics, Le., dynamics in speech can be ob served on, for instance, short-time acoustical, linguistic-segmental, and utterance-wise temporal scales. M 3 is an extension of the Gaus sian mixture model and is considered a hierarchical mixture model, where mixture components in each time scale will change at inter vals of the corresponding time unit. We derive a fully Bayesian treat ment of the multi-scale mixture model based on Gibbs sampling. The advantage of the proposed model is that each speaker cluster can be precisely modeled based on the Gaussian mixture model unlike conventional single-Gaussian based speaker clustering (e.g., using the Bayesian Information Criterion (BIC)). In addition, Gibbs sam pling offers the potential to avoid a serious local optimum problem. Speaker clustering experiments confirmed these advantages and ob tained a significant improvement over the conventional BIC based approaches.

Nested Gibbs sampling for mixture-of-mixture model and its application to speaker clustering

“…A Bayesian approach can make the estimation of mixture-of-mixtures models more robust. For example, maximum a posterior (MAP) and variational Bayes (VB)-based methods have been applied to estimate the mixture of Gaussian mixture models (MoGMMs) [1,18,19]. However, the VB-based approach often still suffers from a large bias when the amount of data is limited [20].…”

“…In [13][14][15], an expectation-maximization (EM) approach [16] was used to estimate mixture-ofmixture models by augmenting observations with two-level (higher-level and lower-level) latent variables. However, this maximum-likelihood-based approach often suffers from an overfitting problem when applied to high-dimensional data [1,17]. A Bayesian approach can make the estimation of mixture-of-mixtures models more robust.…”

“…Thus assume, we that each higher-level observation follows a 1 Department of Communications and Computer Engineering, Waseda University, Tokyo, Japan 2 Mitsubishi Electric Research Laboratories, MA, USA. unique distribution, which represents each speaker's characteristics.…”

“…Lower-level observations correspond to frame-wise observations comprising each utterance, where their variation is caused by the differences in the contents of speech. To cluster utterances by a speaker, we need to derive a suitable mathematical representation of an utterance for extracting each speaker's characteristics independently of the contents of the their speech [1].…”

“…Lower-level observations correspond to frame-wise observations comprising each utterance, where their variation is caused by the differences in the contents of speech. To cluster utterances by a speaker, we need to derive a suitable mathematical representation of an utterance for extracting each speaker's characteristics independently of the contents of the their speech [1].An effective approach for representing higher-level observations is modeling as stochastic distributions. Thus assume, we that each higher-level observation follows a 1 Department of Communications and Computer Engineering, Waseda University, Tokyo, Japan 2 Mitsubishi Electric Research Laboratories, MA, USA.…”