5th International Conference on Spoken Language Processing
Earlier work in parametric modeling of distortions for robust speech recognition has focussed on estimating the distortion parameter using maximum likelihood and other techniques as a point in the parameter space, and treating this estimate as if it is the true value in a plug-in maximum a posteriori(MAP) decoder. This approach is deficient in most real environments where, due to many reasons, the value of the distortion parameter varies significantly. In this paper we introduce an approach which combines the power of parametric transformation and Bayesian prediction to solve this problem. Instead of approximating the distortion parameter with a point estimate, we average over its variation, thus taking into consideration the distribution of the parameter as well. This approach provides more robust performance than the conventional maximum-likelihood approach. It also provides the solution that minimizes the overall error given the distribution of the parameter. We present results to demonstrate the robustness and effectiveness of the predictive approach.
Bibliographic reference. Surendran, Arun C. / Lee, Chin-Hui (1998): "Predictive adaptation and compensation for robust speech recognition", In ICSLP-1998, paper 0859.