Phase AutoCorrelation (PAC) features for noise robust speech recognition

In this paper, we introduce a new class of noise robust features derived from an alternative measure of autocorrelation representing the phase variation of speech signal frame over time. These features, referred to as Phase AutoCorrelation (PAC) features include PAC-spectrum and PAC-MFCC, among others. In traditional autocorrelation, correlation between two time delayed signal vectors is computed as their dot product. Whereas in PAC, angle between the vectors in the signal vector space is used to compute the correlation. PAC features are more noise robust because the angle is typically less affected by noise than the dot product. However, the use of angle as correlation estimate makes the PAC features inferior in clean speech. In this paper, we circumvent this problem by introducing another set of features where complementary information among the PAC features and the traditional features are combined adaptively to retain the best of both. An entropy based feature combination method in a multi-layer perceptron (MLP) based multi-stream framework is used to derive an adaptively combined representation of the component feature streams. An evaluation of the combined features using OGI Numbers95 database and Aurora-2 database under various noise conditions and noise levels show significant improvements in recognition accuracies in clean as well as noisy conditions.

► We propose noise robust features derived from Phase autocorrelation (PAC).
► PAC uses angle between signal frame vectors as a measure of correlation.
► The use of angle improves noise robustness but degrades clean speech recognition.
► An adaptive feature combination approach improves clean speech recognition as well.
► We report experimental evaluations using OGI Numbers95 and Aurora-2 databases.