Research on Speech Separation Method Based on Deep Neural Network
<p>Yandi Luo, Ming Fang</p> · Academic Journal of Computing & Information Science · 2023
This paper focuses on the single-channel speech separation problem and uses deep neural network technology to deeply analyze the loss function, feature extraction and quality evaluation indexes. By improving the existing separation algorithms, a new method combining the jointly constrained loss function and the integrated optimizer is proposed. The study first examines the background and significance of research in the field of speech separation and outlines the current state of research. Then, the basics of speech separation technology and the underlying principles of deep neural networks are introduced in detail. The newly proposed system structure and quality evaluation metrics are used to compare the separation effects of different methods. In addition, the study improves the performance of the algorithm and enhances the ability of the model to avoid local optimums and improve the training efficiency through the proposed joint constrained loss function and integration optimizer.