Research on x-vector speaker recognition algorithm based on Kaldi. (22nd July 2022)

• Record Type: Journal Article
• Authors: Zhao, Hong; Yue, Lupeng; Wang, Weijie; Zeng, Xiangyan
• Is Part Of: International journal of computing science and mathematics. Volume 15:Number 3(2022)
• Page Start: 199
• Page End: 212
• Publication Date: 2022-07-22
• Subjects: spectrogram -- attention mechanism -- x-vector -- speaker recognition -- Kaldi; Mathematics -- Periodicals; Computer science -- Periodicals; Mathematics -- Data processing -- Periodicals; 510.285
• Languages: English
• ISSNs: 1752-5055

Abstract

This paper presents a convolutional neural network with an attention mechanism for analysing the spectrogram in an x-vector based speaker recognition system. First, the convolutional neural network (CNN) is used to extract the features of the spectrogram. Then, an attention mechanism is designed to calculate the frame weight in the statistical pooling layer. Finally, probability linear discriminant analysis (PLDA) is used as a back end classifier. The system is implemented using Kaldi speech recognition tools and tests on the Voxceleb1 database. The experimental results show that the combination of spectrogram and CNN gains a relative improvement of 6.7% in equal error rate (EER) compared with the x-vector baseline system. The attention mechanism for the statistical layer further leads to a relative improvement of 26.1%. Overall the proposed method outperforms state-of-the-art methods on the Voxceleb1 database.