Speaker Recognition Based On Stacked Auto-encoders

Speaker recognition is a biometric identification technique based on the features of one's speech,which will be influenced by not only the physical framework of the vocal character-istic but also the behavioral trait when speaking[34].Various technologies can be used to for speaker recognition purpose such as hidden Markov models,Gaussian mixture models,GSV-SVM,however in reality the performance of the speaker recognition system declines sharply under noisy environment as well as when comes to channel mismatching[29].The state-of-the-art performance was presented by i-vector framework[10]which has become popular in the field of speaker recognition.In recent years,deep learning theory has been successfully applied in many machine learning fields like speech recognition thanks to the strong nonlinear modeling capabilities and feature extraction.For the sake of further improve the performance and robustness of speaker recognition under noisy and cross channel environments,deep learning technic is applied into the modeling framework of speaker recognition in this thesis.We intend to take advantage of the power of abstraction and representation of Stacked Auto-encoders and neural network,coming with the traditional i-vector framework to present a new method for speaker recognition.Also,different variants of auto-encoders will be discussed and explored.