Research On Distance And Similarity Metric Learning For Speaker Recognition

The speaker recognition technology uses speech as the recognition feature and can automatically recognize the identity of the speaker by processing the speaker's speech signal.At present,the speaker recognition system based on the I-vector model has good performance and has become the state-of-the-art in speaker recognition field.This paper mainly studies the speaker recognition algorithm based on I-vector model,and apply the distance and similarity metric learning algorithm to speaker recognition.The metric learning algorithm can decrease the personal intrinsic variances on the system by making the samples of the same class closer and the different class samples farther.This paper mainly studies the application of KISS(Keep It Simple and Straightforward)and SUB-SML(intra-person Subspace Similarity Metric Learning)metric learning algorithms for speaker recognition system.The KISS algorithm has the characteristics of large-scale data sets and scalability,and it can learn metric matrices from equivalence constraints.Moreover,the learned metric matrices can guarantee the strong correlation of similar samples.In this paper,the KISS metric learning algorithm is applied to the speaker recognition system,and the KISS metric matrix linear transformation is performed on the i-vector to decrease intra-class variance.The experiments were run on the NIST SRE 2014 i-vector challenge database,and cosine similar classifier and Mahalanobis distance classifier are used for scoring.The experimental results show that cosine similar classifier can get better performance when using KISS metric learning,and the Mahalanobis distance classifier can get best performance.In addition,Mahalanobis distance classifier with KISS metric learning and PLDA can be fused at the score level,and the results show that the performance was further improved.SUB-SML algorithm formulate the learning objective by considering both the robustness to the large intrapersonal variations and the discrimination for separating similar i-vectors from dissimilar i-vectors.The algorithm combines the bilinear similarity function and the Mahalanobis distance and propose a generalized similarity metric to measure the similarity of an i-vector pair.The NIST SRE 2014 i-vector challenge only provides i-vector label information,so it is necessary to construct an ivector pair set for the SUB-SML algorithm.In this paper,we propose a new method to construct i-vector pairs which is based on the Euclidean distance.In this paper,the SUB-SML metric learning algorithm is applied to the speaker recognition system.The new method used to construct the training sample pair set is performed on the SUBSML algorithm,and the SUB-SML metric matrix linear transformation is performed on the i-vector.The experiment results show that the performance of the SUB-SML model is better than that of the cosine similar classifier,and the performance of the SUB-SML model with the i-vector pairs by the new method is significantly improved.The SUB-SML algorithm combines the Mahalanobis distance and the bilinear similarity function.If only the Mahalanobis distance measure or bilinear similarity measure is considered,the SUB-ML and SUB-SL algorithms are obtained.In this paper,SUB-ML and SUB-SL metric learning algorithms are applied to the speaker recognition system respectively.The specific experimental part refers to the SUB-SML algorithm.Experiments show that the learning algorithm based on SUB-ML and SUB-SL can improve speaker system recognition efficiency.In addition,the SUB-ML model and the SUB-SL model can be fused at the score level,and the results show that the performance was further improved.