A study on bias-based speech signal conditioning techniques for improving the robustness of automatic speech recognition

Automatic speech recognition (ASR) performs poorly when the training conditions greatly mismatch the testing conditions. Additive background noises and channel distortion are responsible mostly for these mismatches. These mismatches introduce highly non-linear terms in the acoustic model of speech in both log-spectral and the cepstral domains. Current ASR is based on simple linear approximation of these non-linear functions in the cepstral domain in order to avoid model complexities. This linear modeling approach transforms the channel distortion into an additive bias term in the cepstral domain under the assumption of high SNR, which is barely true in practical situations. Several algorithms have been developed to estimate this bias term and make compensations either in feature space or in the model domain to improve the robustness of ASR. In this paper, we explore these bias estimate techniques for both stationary and non-stationary acoustic environments to find their applicability for self-adaptable ASR.