Text Independent Speaker Recognition Based On Deep Learning Framework

Speaker recognition refers to the use of speech signals to identify the speaker.It can be widely used in national defense security,public security,finance,social security and intelligent terminals etc.,therefore it has important research significance.In this paper,we focus on the research of text-independent speaker identification,and have proposed two text-independent speaker identification algorithms.One is a text-independent speaker identification algorithm through combining timbre features with gender and accent auxiliary features,and the other is a text-independent speaker identification algorithm based on multi-task learning.The main contributions and innovations of our work include:(1)A text-independent speaker identification algorithm which combines the timbre features with the gender and accent auxiliary features for classification is proposed.Currently,most speaker identification algorithms only use timbre features for classification,but some studies have pointed out that the robustness of such kind of features to audio degradation is poor,and then only using such kind of features for classification will limit the recognition performance.In addition,when it is difficult for human to recognize the speaker,in addition to the timbre features,some other information of the speaker is usually mined as an aid for recognition.Based on the above discussion,in this paper,we propose to introduce the speaker's gender and accent information as the auxiliary features,and propose to combine timbre features with gender and accent auxiliary features for classification.Based on the attention mechanism,an attentional network is designed to generate auxiliary features by embedding gender and accent information into the timbre features.For the proposed algorithm,it is the first time that the gender and accent information of the speaker is jointly mined and combined with timbre features for classification.Experimental results have confirmed the correctness of the proposed idea,that is,in addition to the commonly used timbre features,the introduction of speakers' other information will help to improve the recognition performance.(2)A text-independent speaker identification algorithm based on multi-task learning is proposed.One of the key problems in speaker identification technology is how to extract speakers' specific features for recognition.If the speech features that reflect the unique characteristics of different speakers can be extracted,they will help to improve the recognition performance.Some research has pointed out that the speech features extracted by speaker verification method can well reflect the characteristics of the speaker.Inspired by it,in this paper we have proposed a speaker identification algorithm based on multi-task learning.In this method,a network model is built to realize speaker verification and speaker identification at the same time.The speaker verification network is used to assist speaker identification network to extract the speech features that can reflect speaker specific characteristics,so as to improve the speaker identification accuracy.