Enhancing Pitch Robustness of Speech Recognition System through Spectral Smoothing

摘要

In this paper, we present a novel approach for front-end speech parameterization that is more robust towards pitch variations than the most commonly used technique. Earlier works have shown that, insufficient smoothing of magnitude spectrum leads to pitch-induced distortions. This, in turn, results in poor performance of speech recognition system especially for high-pitched child speakers. To overcome this shortcoming, the short-time magnitude spectrum is first decomposed into several components using a modified version of empirical mode decomposition (EMD). Next, the lowest-order component is discarded and the spectrum is reconstructed using the rest of the higherorder modes for sufficiently smoothing the spectrum. The Mel-frequency cepstral coefficients (MFCC) are then extracted using the smoothed spectra. The signal domain analyses presented in this paper demonstrate that the ill-effects of pitch variations get significantly reduced by the inclusion of proposed spectral smoothing module. In order to statistically validate the same, an automatic speech recognition system is developed using speech data from adult speakers. To simulate large pitch differences, evaluations are performed on a test set which consists of speech data from child speakers. Inclusion of proposed spectral smoothing module leads to a relative improvement of 12% over the baseline system employing acoustic modeling based on deep neural network.