Boosting DNN-based speech enhancement via explicit transformations

摘要

In this study, we investigate on the learning behaviors of DNN by explicit feature transformations. As a demonstration, linear and logarithm transformations, corresponding to the amplitude spectra and log-power spectra, are compared with the same minimum mean squared error (MMSE) objective function for optimizing DNN parameters. Based on the experimental analysis of the DNN learning behaviors, we make an interesting observation that the learning with the amplitude spectra tends to improve the speech intelligibility while the learning with the log-power spectra yields better speech quality. By leveraging on this strong complementarity, the feature concatenation with two transformations for the input layer and post-processing with two learned targets are proposed to boost DNN-based speech enhancement.