Constrained adaptation of histogram equalization for robust speech recognition

Parametric histogram equalization (PHEQ) is a simple and effective way to reduce noise and channel effects in automatic speech recognition. Recently, we proposed to adapt the PHEQ transform towards individual test utterance to further improve the performance of PHEQ. In this paper, we improve the adaptation of PHEQ by exploiting the correlation between speech feature elements. Previously, PHEQ is applied to each feature element independently. In the case of Mel-frequency cepstral features (MFCC), this may not be optimal as it ignores the fact that the delta and acceleration features are derived from the static features and they are correlated. In this paper, we proposed to maintain the relationship between static, delta, and acceleration features to impose a constraint in PHEQ adaptation such that the PHEQ parameters will be more reliably estimated for short utterances. Experiments on both Aurora-2 and Aurora-4 tasks confirmed that it is beneficial to maintain the relationship between feature elements especially for short sentences.