PHASEN: A Phase-and-Harmonics-Aware Speech Enhancement Network

• DOI: 10.1609/aaai.v34i05.6489
• Publication Date: 2020-06-17T08:27:54Z
• Authors: Dacheng Yin; Chong Luo; Zhiwei Xiong; Wenjun Zeng
• Keywords: Spectrogram; Harmonic

Time-frequency (T-F) domain masking is a mainstream approach for single-channel speech enhancement. Recently, focuses have been put to phase prediction in addition amplitude prediction. In this paper, we propose phase-and-harmonics-aware deep neural network (DNN), named PHASEN, task. Unlike previous methods which directly use complex ideal ratio mask supervise the DNN learning, design two-stream network, where stream and are dedicated We discover that two streams should communicate with each other, crucial addition, frequency transformation blocks catch long-range correlations along axis. Visualization shows learned matrix implicitly captures harmonic correlation, has proven be helpful T-F spectrogram reconstruction. With these innovations, PHASEN acquires ability handle detailed patterns utilize patterns, getting 1.76dB SDR improvement on AVSpeech + AudioSet dataset. It also achieves significant gains over Google's On Voice Bank DEMAND dataset, outperforms by large margin four metrics.