Weighted Pairwise Constraints Metric Learning Algorithm In Speaker Recognition

Speaker recognition is a technology that can identify the speaker by processing and analyzing the utterance of speaker. At present, speaker recognition has a very wide range of uses as an intelligent interface of human and computer. It has many convenient, economical, reliable and good extensible advantages, and has been widely used in remote authentication, military and medical fields, especially the application of identity recognition based on Telecom network, the prospect of speaker recognition technology is also very broad.How to measure the similarity between the samples of the speaker's speech effectively is one of the key problems in the field of speaker recognition. In the field of pattern recognition, there are a lot of methods measuring the similarity of samples, and the commonly used methods are distance measuring methods, such as cosine distance scoring and Mahalanobis distance scoring. The art of date method based on I-vector speaker recognition system commonly used cosine distance to measure the similarity between the speaker, and cosine distance scoring method is according to the angle between the vectors to judge the similarity between samples, which is difficult to measure the difference in the amount of samples. According to the Euclidean distance between the new samples in the sample space which can reflect the similarity of sample space, the method of Mahalanobis distance can measure the similarity between samples. The mapping matrix, which is used to project the sample space, is called the metric matrix, which can be used to describe the similarity of the sample space. The sample spaces described by different matrix are different also, and only the correct measurement matrix can reflect the similarity between samples. Metric learning is based on the information of the training samples, and a distance measure matrix is obtained by automatic learning, which is usually used to calculate the distance between the target samples and the similarity of unknown data.The main work and innovation of this paper are as follows:(1) This paper study the sub algorithm(subspace metric learning, SUB-ML) of subspace similarity metric learning algorithm(SUB-ML) proposed by Cao et al. According to the influence of weighted constraint of the similarity of training samples and dissimilar training samples of metric learning training process, the weighted pairwise constraint metric learning(WPCML) algorithm is proposed. The In this algorithm, a metric matrix for describing the similarity of samples is trained by the constraint information of pairwise training samples, which is used to calculate the distance between samples. The algorithm is simple and effective, and there is a global optimal solution, which can quickly obtain the distance measure matrix which satisfies the condition. The United States National Institute of standards and Technology(NIST) speaker recognition evaluation(SRE) 2008 data set of experimental results show that the performance of WPCML algorithm used to train the metric matrix for Mahalanobis distance scoring is better than that of cosine distance scoring.(2) The construction of training samples is an important problem in metric learning. Appropriate training samples can guide the training process more effectively. Most of the metric learning algorithms use the randomly construction method to construct the metric learning training sample pair set. In this paper, the construction method of training samples is studied, and the method of selecting training samples is proposed to construct the matrix. The method of choosing the training samples to construct training data set can improve the performance of the system further, and the performance is better than the most popular PLDA classifier.