Multi-Metrics Learning for Speech Enhancement

This paper presents a novel deep neural network (DNN) based speech enhancement method that aims to enhance magnitude and phase components of speech signals simultaneously. The novelty of the proposed method is two-fold. First, to avoid the difficulty of direct clean phase estimation, the proposed algorithm adopts real and imaginary (RI) spectrograms to prepare both input and output features. In this way, the clean phase spectrograms can be effectively estimated from the enhanced RI spectrograms. Second, based on the RI spectro-grams, a multi-metrics learning (MML) criterion is derived to optimize multiple objective metrics simultaneously. Different from the concept of multi-task learning that incorporates heterogeneous features in the output layers, the MML criterion uses an objective function that considers different representations of speech signals (RI spectrograms, log power spectrograms, and waveform) during the enhancement process. Experimental results show that the proposed method can notably outperform the conventional DNN-based speech enhancement system that enhances the magnitude spectrogram alone. Furthermore, the MML criterion can further improve some objective metrics without trading off other objective metric scores.