A novel speech activity detection algorithm based on the fusion of time domain and frequency domain features
LIU Huan;WANG Jun;LIN Qiguang;WANG Shitong;School of Digital Media,Jiangnan University;Baihu Technology Company of Wuxi;
In order to improve the adaptability and robustness of speech activity detection,a novel algorithm for speech activity detection(SAD) is proposed based on the integration of time domain and frequency domain features. In the proposed method,three features,i. e. harmonicity,clarity,periodicity are extracted and combined together with principal component analysis. The candidates of the endpoints are detected by double-threshold method. SVM is utilized to determine the final set of endpoints based on the candidates. Experimental results indicate that the proposed SAD method is effective and provides superior and consistent performance across various noise and distortion levels.