Study On Deep Learning-based Speaker Recognition

Speaker recognition is called voiceprint identification. It is a kind of authentication technology. Speaker recognition technology has many advantages, including high user acceptance, low equipment costs, strong scalability and easy to transplantation. It is widely used in military field, bank system, internet security and judicial security. Speaker recognition technology is related to our life closely and has great research value and practicality.This thesis mainly studies the speaker recognition system with deep learning model. Some basic system performance testing is completed and discussed, and this paper modified speech feature parameters and statistical method to obtain a higher speaker recognition system rate. What are this paper talking about is as follows:(1)The basic performance of system based on deep learning. The deep learning model is introduced in speaker recognition system. The impact of the different length of speech units on speaker recognition system rate is studied. On the same test condition, the impact of different speech features on speaker recognition system rate is also studied. The impact of different layers and nodes of deep learning model on system recognition rate is studied. The accuracy and reliability of deep learning model applied on speaker recognition system is proved.(2)Based on human auditory characteristics, we apply a new speech feature by combining MFCC with GFCC to speaker recognition system to improve the recognition rate.(3)Considering the traditional system statistics algorithm for multi-speaker recognition leads to misjudgment, we proposed a modified statistics algorithm for multi-speaker recognition system. The effectiveness of modified method is proved by experiments.