A Hybrid CNN-Transformer Architecture for Noise Cancellation in Complex Environments

This paper proposes a speech enhancement algorithm that fuses a convolutional neural network (CNN) with a Transformer. It aims to effectively suppress background interference in complex noise environments and improve the clarity and intelligibility of speech signals. A weighted source-to-distortion ratio (wSDR) loss function is used to obtain the training gradient of the balanced model against the speech and noise signals by adjusting the parameter alpha. The results of the verification on the AISHELL-1 dataset show that the CNN+Transformer model has achieved an average PESQ score of 2.518 and an STOI score of 0.7234 under most test conditions.