Band-Split Inter-SubNet: Band-Split with Subband Interaction for Monaural Speech Enhancement

Speech enhancement models are developed to improve quality and intelligibility of speech for numerous daily applications. With the rapid development of technology, the neural network based speech enhancement models show significantly improved performance. The subband-based models focus on local spectral patterns and achieve outstanding results with fewer parameters. In this paper, we propose a subband-based composite model named Band-Split Inter-SubNet. It adopts the new constant-Q band-split setting to mimic human auditory perception. The proposed model demonstrates superior performance to other state-of-the-art models on the DNS Challenge - Interspeech 2021 dataset. Detailed analyses on experimental results demonstrate that the proposed band-split setting is effective, and the influence of neighboring frequency bins on the center-frequency bin across different frequency bands varies slightly.