SETransformer: Speech Enhancement Transformer

Speech enhancement is a fundamental way to improve speech perception quality in adverse environment where the received speech is seriously corrupted by noise. In this paper, we propose a cognitive computing based speech enhancement model termed SETransformer which can improve the speech quality in unkown noisy environments. The proposed SETransformer takes advantages of LSTM and multi-head attention mechanism, both of which are inspired by the auditory perception principle of human beings. Specifically, the SETransformer pocesses the ability of characterizing the local structure implicated in the speech spectrum and has more lower computation complexity due to its distinctive parallelization perfermance. Experimental results show that, compared with the standard Transformer and the LSTM model, the proposed SETransformer model can consistently achieve better denoising performance in terms of speech quality (PESQ) and speech intelligibility (STOI) under unseen noise conditions.