Hybrid Particle Swarm Optimization-Deep Neural Network Model for Speaker Recognition
V. Sesha Srinivas, Santhirani Ch, Tengyue Bian, Fangzhou Chen, Li Xu, Jess Villalba, Nanxin Chen, David Snyder, Daniel Garcia-Romero, Najim Dehak, Shuping Peng, Tao Lv, Xiyu Han, Shisong Wu, Heyong Zhang, Emma Jokinen, Rahim Saeidi, Tomi Kinnunen, Paavo Alku, Michael Jessen, Jakub Bortlk, Petr Schwarz, Yosef Solewicz, Ville Vestman, Dhananjaya Gowda, Md Sahidullah, Paavo Alku, Tomi Kinnunen, Ing-Jr Ding, Jia-Yi Shi, Johan Rohdin, Anna Silnova, Mireia Diez, Oldich Plchot, Ondej Glembek, V Ramaiah, R Rao, P Valsalan, Manimegalai, S Augustine, V Vrabie, P Granjon, C Serviere, J Morales-Cordovilla, A Peinado, V Snchez, J Gonzlez, Manoela Kohler, Marley Vellasco, Ricardo Tanscheit, J Ren, S Yang, Ravi Kumar Vuddagiri, Hari Vydana, Anil Vuppala
Abstract
Nowadays, speaker recognition is considered as a current research topic. Moreover, the voice biometrics which is attained from the speaker's behavior or physical related features offers a pattern of data that contains sensitive information regarding the speaker. The efficiency of speaker recognition systems is observed to minimize expeditiously because of the mismatch incidence, such as noise and channel degradations. With the aspire to promise security and effectual recognition, a Hybrid Particle Swarm Optimization-based Deep Neural Network (Hybrid PSO-based DNN classifier) is used to identify a speaker for that the frequency-dependent features, like MKMFCC, autocorrelation, and spectral skewness, are exploited. The classification is done by exploiting the DNN classifier based on feature extraction and classifier is performed optimally by exploiting the proposed Particle Swarm Optimization. Finally, the simulation analysis of the proposed technique is compared with the LM, SVM, GMM, and BSW. It shows the performance of the proposed technique outperforms the conventional techniques concerning the accuracy, FAR FRR.