Improved robust features for speech recognition by integrating time-frequency principal components (TFPC) and histogram equalization (HEQ)

Robustness for speech recognition technologies with respect to adverse environments has been a key issue for real applications. Time-frequency principal components (TFPC) features have been shown to be a set of powerful data-driven features under matched circumstances, while histogram equalization (HEQ) has been proposed as an efficient feature transformation approach to reduce the mismatch between training and testing conditions. It is proposed that TFPC features can be well integrated with HEQ. HEQ generates a well-matched environment, in which TFPC features can be properly utilized. Extensive experiments with respect to the AURORA2 database verified that improved performance in adverse circumstances can be achieved.