Real and imaginary part interaction network for monaural speech enhancement and de-reverberation
Zehua Zhang, Changjun He, Shiyun Xu, Mingjiang Wang · 2023
Speech enhancement and de-reverberation are the key points of speech signal front-end processing. Direct mapping of speech spectrum is a standard method in speech enhancement and de-reverberation. In previous studies, the real and imaginary parts of the spectrum are estimated separately, which would make the error of direct estimation more significant. This paper proposes a real and imaginary parts interactive network (RIINet) in which the branches of real and imaginary parts interact through a linear complex attention mechanism. In addition, this paper proposes a time-frequency analysis module to improve the performance of local time-frequency feature extraction effectively. RIINet achieved state-of-the-art performance in the Deep Noise Suppression Challenge using minimal model parameters. Excellent performance has also been achieved in the noise-reverberation speech enhancement test set.