Uncertainty Decoding for Noise Robust Automatic Speech Recognition

This report presents uncertainty decoding as a method for robust automatic speech recognition for the Noise Robust Automatic Speech Recognition project funded by Toshiba Research Europe Limited. The effects of noise on speech recognition are reviewed and a general framework for noise robust speech recognition introduced. Common and related noise robustness techniques are described in the context of this framework. Uncertainty decoding is also presented in this framework with the goal of providing fast noise compensation through the propagation of uncertainty to the de-coder. Two forms are discussed, the Joint and SPLICE methods, and evaluated on the medium vocabulary Resource Management corpus at a range of artificially produced noise levels. It was found that the uncertainty decoding algorithms did not meet the performance of a matched system, but were more accurate than the baseline SPLICE enhancement technique and low numbers of CMLLR transforms. Notation A matrix of arbitrary dimensions A T transpose of matrix A |A| determinant of matrix A A-1 inverse of matrix A I identity matrix x scalar quantityˆx estimate of the true value of x x vector of arbitrary dimensions T number of frames in a sequence of observations M number of GMM components in the acoustic model