Spatial histogram equalization of complex-valued acoustic spectra in modulation domain for noise-robust speech recognition

摘要

This paper proposes to enhance the complex-valued acoustic spectrograms of speech signals via the technique of histogram equalization (HEQ) to produce noise-robust features for recognition. The presented method extends our previous work in the task of spectrogram enhancement and has two significant aspects. First, we process the real and imaginary parts of acoustic spectrograms separately, and therefore both of the corresponding magnitude and phase components can be enhanced implicitly. Second, we apply FIR filters to the intra-frame acoustic spectra to acquire the respective local structural statistics, which are subsequently employed to perform various types of HEQ on the acoustic spectrograms for robustifying the resulting speech features. All experiments were carried out on the Aurora-2 database and task. The performance of the presented methods was thoroughly tested and verified by comparisons with other well-known robustness methods, which reveals the capability of our methods in promoting the noise robustness of speech features.