Design And Research In Speaker Recognition System

In recent years,the rapid development of machine learning is deeply affecting people's daily life,of which human-computer interaction is the focus of machine learning.In the implementation of human-computer interaction,the user identity needs to be determined and then the service will be provided accordingly.Using the voice information to achieve the identification is a user-friendly option,which motivates the research in speaker recognition task.In this paper,we focus on the topic of constructing a robust and accurate speaker recognition system and propose two sorts of speaker recognition systems.The first sort is a speaker recognition system based on online i-vector.This system combines the advantages of i-vector model and GMM-UBM system,which achieves a better result than the i-vector-PLDA baseline system on text-dependent speaker recognition tasks.The second sort systems are end-to-end speaker recognition systems based on neural networks,which achieves the effect of the overall optimization during training.One of the end-to-end systems,which is based on the Triplet loss function,achieves better recognition performance than the i-vector-PLDA baseline system on the short-term text-independent speaker recognition task.