Approximate Bayesian Inference for Robust Speech Processing

Abstract

Speech processing applications such as speech enhancement and speaker identification rely on the estimation of relevant parameters from the speech signal. These parameters must often be estimated from noisy observations since speech signals are rarely obtained in ‘clean’ acoustic environments in the real world. As a result, the parameter estimation algorithms we employ must be robust to environmental factors such as additive noise and reverberation. In this work we derive and evaluate approximate Bayesian algorithms for the following speech processing tasks: 1) speech enhancement 2) speaker identification 3) speaker verification and 4) voice activity detection. Building on previous work in the field of statistical model based speech enhancement, we derive speech enhancement algorithms that rely on speaker dependent priors over linear prediction parameters. These speaker dependent priors allow us to handle speech enhancement and speaker identification in a joint framework. Furthermore, we show how these priors allow voice activity detection to be performed in a robust manner. We also develop algorithms in the log spectral domain with applications in robust speaker verification. The use of speaker dependent priors in the log spectral domain is shown to improve equal error rates in noisy environments and to compensate for mismatch between training and testing conditions.