Bayesian approaches to uncertainty in speech processing

Many techniques in speech processing require inference based on observations that are often noisy, incomplete or scarce. In such situations, it is necessary to draw on statistical techniques that themselves must be robust to the nature of the observations. The Bayesian method is a school of thought within statistics that provides such a robust framework for handling “difficult” data. In particular, it provides means to handle situations where data are scarce or even missing. Three broad situations are outlined in which the Bayesian technique is helpful to solve the associated problems. The analysis covers eight publications that appeared between 1996 and 2011. Dialogue act recognition is the inference of dialogue acts or moves from words spoken in a conversation. A technique is presented based on counting words. It is formulated to be robust to scarce words, and extended such that only discriminative words need be considered. A method of incorporating formant measurements into a hidden Markov model for automatic speech recognition is then outlined. In this case, the Bayesian method leads to a re-interpretation of the formant confidence as the variance of a probability density function describing the location of a formant. Finally, the Gaussian model of speech in noise is examined leading to improved methods for voice activity detection and for noise robustness.