Loudness pattern-based speech quality evaluation using Bayesian modeling and Markov chain Monte Carlo methods

Abstract

This work presents a speech quality evaluation method which is based on Moore and Glasberg's loudness model and Bayesian modeling. In the proposed method, the differences between the loudness patterns of the original and processed speech signals are employed as the observed features for representing speech quality, a Bayesian learning model is exploited as the cognitive model which maps the features into quality scores, and Markov chain Monte Carlo methods are used for the Bayesian computation. The performance of the proposed method was demonstrated through comparisons with the state-of-the-art speech quality evaluation standard, ITU-T P.862, using seven ITU subjective quality databases.