Histogram equalization of real and imaginary modulation spectra for noise-robust speech recognition

Histogram equalization (HEQ) of acoustic features has received considerable attention in the area of robust speech recognition because of its relative simplicity and good empirical performance. This paper presents a novel HEQbased feature extraction approach that performs equalization in both acoustic frequency and modulation frequency domains for obtaining better noise-robust features. In particular, the real and imaginary acoustic spectra are first individually transformed to the modulation domain via discrete Fourier transform (DFT). The HEQ process is then carried on the corresponding magnitude modulation spectra so as to compensate for the noise distortions. Finally, the equalized modulation spectra are converted back to form the real and imaginary acoustic spectra, respectively. By doing so, we can enhance not only the magnitude but also the phase components of the acoustic spectra, and thereby create more noise-robust cepstral features. The experiments conducted on the Aurora-2 clean-condition database and task reveal that the presented approach delivers superior recognition accuracy in comparison with some other HEQ-related methods and the well-known advanced front-end (AFE) extraction scheme, which supports the potential utility of this novel approach.