Leveraging threshold denoising on DCT-based modulation spectrum for noise robust speech recognition

This paper presents a novel noise robustness algorithm to enhance speech features in noisy speech recognition. In the presented algorithm, the temporal speech feature sequence is first converted to its spectrum via discrete cosine transform (DCT), and then the DCT spectrum is compensated by a thresholding function in order to further shrink the smaller portion. Finally, the updated DCT spectrum is converted back to the temporal domain to obtain the new feature sequence. One advantage of the presented method is that the overall compensation process is unsupervised in the sense that no information about noise embedded in speech signals is required. The evaluation via the Aurora-2 connected digit database and task revealed that the presented method can provide significant improvement in recognition accuracy to the speech features pre-processed by any of the statistics normalization algorithms, including cepstral mean and variance normalization (CMVN), MVN plus ARMA filtering (MVA) and cepstral gain normalization (CGN). We further showed that, using the presented method, simply compensating the low frequency portion gives similar performance on a par with that achieved by compensation over the entire frequency band.