ISCA Archive Eurospeech 1995
ISCA Archive Eurospeech 1995

The use of maximum a posteriori parameters in linear prediction of speech

G. M. K. Saleh, M. Niranjan, W. J. Fitzgerald

We present an approach to linear prediction parameter estimation and model order selection that utilises Bayesian inference. The addition of a penalty term, or regulariser, to the conventional linear prediction data error term prior to minimising it facilitates the estimation of the maximum a posteriori parameters. A direct equivalence can be drawn between the type of regulariser used and the prior assumptions regarding the solution to a linear prediction problem. Mackay's Bayesian Evidence framework is used for the estimation of linear prediction parameters that reflect the role that prior assumptions play during the analysis of a speech segment. Quadratic regularisers are utilised to parametrise speech signals and the results are demonstrated with formant tracking and analysis-synthesis examples.