Exploring Feature Enhancement in the Modulation Spectrum Domain via Ideal Ratio Mask for Robust Speech Recognition

Development of robustness techniques is of paramount importance to the success of automatic speech recognition (ASR) systems. In this paper, we present a novel use of the ideal ratio mask (IRM) method to improve ASR robustness. IRM was originally proposed for time-frequency (T-F) masking-based speech enhancement and has shown considerable promise in preserving the intelligibility of a noisy mixture signal. Further, IRM is alternatively used to normalize the intermediate representations of speech feature vector sequences, in a holistic manner, for both training and test utterances. Finally, we instead treat IRM as a data augmentation method, conducted on speech feature vectors of training utterances or their intermediate representations, to generate additional augmented data for increasing the diversity of training data. A series of experiments carried out on the standard Aurora-4 database and task confirm the effectiveness of our methods.