In this paper, a novel method is proposed to enhance the complex-valued acoustic spectrograms of speech signals via replacing the magnitude part of the corresponding modulation spectrum in order to create noise-robust feature representation for recognition. All the evaluation experiments implemented on the Aurora-2 digit database and task show that the presented method performs better than the baseline MFCC and several well-known noise-robust techniques. These results apparently reveal that this novel method alleviates the effect of noise in speech features significantly.
展开▼