A time-synchronous histogram equalization for noise robust speech recognition

Abstract

The histogram equation (HEQ) technique is commonly adopted for feature space normalization in speech recognition systems. In this technique, a transform function is calculated directly from the histograms of both training and test data, and the nonlinear effects of additive noise are compensated. In order to estimate the transform function accurately, a certain amount of data are required. However, this is not suitable for real-time application because at least several seconds of evaluation data need to be accumulated before the transform function can be calculated. This means that the system cannot start the recognition process until the end of utterance. In this research, we aim to develop a new speech recognition method based on the HEQ technique for real-time processing. This method is called “time-synchronous frame-weighted HEQ (ts-FHEQ).” In the time-synchronous decoding, lack of data for estimating the histogram becomes a major problem. To resolve this problem, we introduce a frame weighting approach, where the degree of transform is controlled according to the number of data frames. Our speech recognition experiments verified that the proposed technique shows good performance and achieves substantial reduction of calculation time.