This paper presents an algorithm for noise compensation in the log-spectral domain. The idea is based on the use of accurate approximations which allow theoretical derivations of the noisy speech statistics, and using these statistics to define a compensation algorithm under a Gaussian mixture model assumption. The algorithm is tested on a digit data base recorded in the car, the word recognition accuracies for the baseline (uncompensated), first order VTS, the proposed method, and the matched test, are 85.8%, 90.6%, 93.1%, and 93.9% respectively. This clearly indicates the performance gain due to the proposed technique.
Download Full PDF Version (Non-Commercial Use)