Low-Delay Speech Enhancement Using Perceptually Motivated Target and Loss

Speech enhancement approaches based on deep neural network have outperformed the traditional signal processing methods. This paper presents a low-delay speech enhancement method that employs a new perceptually motivated training target and loss function. The proposed approach can achieve similar speech enhancement performance compared to the state-of-the-art approaches, but with significantly less latency and computational complexities. Judged by the MOS tests conducted by the INTERSPEECH 2021 Deep Noise Suppression Challenge organizer, the proposed method is ranked the 2nd place for Background Noise MOS, and the 6th place for overall MOS.