Speech Processing Using Bayesian Inference

Abstract

In this paper, we report on the application of Bayesian methods to the analysis of speech signals. Voiced speech can be modelled as a superposition of decaying sinusoids and estimates of the resonant frequencies, decay rates, phases, amplitudes as well as the number of model functions are calculated. The motivation for this model is that in speech analysis, the frequencies and decay rates correspond to formants and bandwidths which are perceptually significant parameters. Speech parameters are estimated by calculating the posterior probabilities for the model parameters, after various nuisance parameters have been marginalised, and it is shown how model order evidence can be calculated. Comparisons with methods such as Minimum Description Length and the Akaike Information Criteria will be made.