The Design And Implementation Of The Fast Search Mechanism In Spekaer Recognition

Although the speaker recognition system has achieved satisfactory results in the aspect of recognition accuracy,with the increasing population,the traditional GMM-based speaker recognition system has to spend longer time to find out the claimant speaker.And its lower recognition efficiency has limited its application to real life in some ways.It was found out that when test uttrances were scored against the speaker models,only a handful of speakers got the higher scores.This means that these speaker models are similar.If we can gather these models together into a cluster via some clustering algorithm,then during the test stage,what we need to do is find out the cluster these models belong to and search out the claimant speaker with really less time campared with the traditioanal GMM-based method.For this,the paper proposed a speaker recognition system based on speaker model clustering algorithm,which can improve the recognition efficiency to some extent without reducing the recognition performance.The traditional speaker model clustering which used K-means algorithm and the approximated KL distance metric to cluster the speaker models has improved the system performance to a large extent.However,this traditional clustering method is sensitive to the initial clustering centres.For this,the paper puts forward an initial clustering algorithm based on a distance matrix acquired by computing the mutual distances between speakers.This method can reasonably selected out the initial centres according to data distributions in speaker space.Then in the test stage,the test uttrance should firstly find out which cluster its claimant speaker belongs to by scoring against the cluster representative and then search out the claimant speaker in the cluster whose respresentatives got the highest scores.When applying the proposed cluster method to the baseline system,we can find that the method can improve the recognition efficiency to a certain degree and reduce the recognition accuracy in the tolerable range.