Leveraging distributional characteristics of modulation spectra for robust speech recognition

Modulation spectrum processing of speech features has recently become an active area of intensive research in the speech recognition community. As for normalization of modulation spectra, spectral histogram equalization (SHE) seems to be one of the most effective techniques that have been used to compensate the nonlinear distortion. In this paper, we investigate a novel use of polynomial-fitting techniques for modulation histogram equalization, which has the advantages of lower storage and time consumption when compared with the conventional SHE methods. Further, we also investigated the possibility of combining our approach with other temporal feature normalization methods. The automatic speech recognition (ASR) experiments were carried out on the Aurora-2 standard noise-robust ASR task. The performance of the proposed approach was thoroughly tested and verified by comparisons with the other popular modulation spectrum normalization methods, which suggests the utility of the proposed approach.