Feature Compensation with Class-based Histogram Equalization for Robust Speech Recognition

In this paper, a new method based on the class-based histogram equalization to compensate the acoustic mismatch between training and test conditions of speech recognizers is proposed. The proposed method improves the speech recognition accuracy in noisy environments by reducing two limitations of the conventional histogram equalization: The discrepancy of phonetic class distributions between training and test speech data, and the non-monotonic transformation caused by the acoustic mismatch in the histogram equalization-based feature domain. The algorithm employs multiple class-specific reference and test cumulative distribution functions, classifies the feature vector for each frame into its corresponding class using the k-means clustering method, and equalizes each feature coefficient by using the corresponding class reference and test distributions. The experiments on the Aurora 2 task proved the effectiveness of the proposed method by reducing averaged error rates by 19% over the conventional histogram equalization method and by 60% over the mel-cepstral-based features, respectively.