부산대학교 도서관

Noise is a critical factor that impacts speech quality. Speech signals are often disturbed by noise during acquisition and transmission. Speech enhancement technology represents an important method to eliminate noise from noisy signal and extract clean speech, which holds significant value in the field of speech recognition, speech communication. This study first categorizes the prior characteristics of different types of noise, followed by an exposition and implementation of representative speech enhancement methods, including traditional signal processing methods and deep learning algorithms. Through three metrics of evaluation, the performance of various speech enhancement algorithms in distinct noise environments is compared, and the Entropy Weight-Topsis (EW-Topsis) method is used for synthesis. Finally, the applicability and development trends of existing speech enhancement algorithms were discussed. The innovation of this study lies in the comparison of different algorithms in various noisy environments using different evaluation metrics. Previous research has mainly focused on two of the evaluation, algorithms and noise. This study presents a comprehensive comparison to synthesize three aspects, providing a unified evaluation reference.