Quantile based histogram equalization for noise robust speech recognition

Abstract

In many practical applications automatic speech recognition systems have to work in adverse acoustic environment conditions. Automatic systems are much more sensitive to the variabilities of the acoustic signal than humans. Whenever noise causes a mismatch between the distribution of the training data and the data that is to be recognized, the recognition word error rates will increase. Quantile based histogram equalization is a method to increase the noise robustness. During the feature extraction it reduces an eventual mismatch between the recognition and training data distributions with a non-linear parametric transformation function. This work describes the algorithm and presents detailed experimental evaluations. Based on the quantiles of the cumulative distributions, the parameters of the transformation functions can be reliably estimated from small amounts of data. The approach is integrated into a modified Mel cepstrum feature extraction, in which the logarithm is replaced by a root function to further increase the noise robustness. The actual transformation that is proposed in this work consists of two steps. First, a power function transformation is applied to each output of the Mel-scaled filter-bank, then neighboring filter are channels combined linearly. To investigate the genericity of the approach and the proposed setup experimental evaluations have been carried out with different speech recognition systems, on several databases with different levels of complexity, ranging from digit strings (SpeechDat Car) to larger vocabulary isolated word (Car Navigation) and continuous speech recognition tasks (Wall Street Journal with added noise). Consistent recognition results were observed on all databases. The modified feature extraction, with the root instead of the logarithm, already outperformed the original baseline on noisy data. Filter channel specific quantile equalization always improved these results, yielding relative improvements between of 5% and 50%, depending on the recognition task and the mismatch of the data. Finally, the combination of neighboring filter channels was able to reduce the error rates somewhat further, especially if the noise, like car noise, was band limited.