Feature vector classification based on likelihood ratio for speaker identification

Abstract

This paper describes a new feature vector classification method for speaker identification. Speaker identification system selects the speaker who has the highest likelihood when the test utterance is given. Speaker identification system is composed of two distinct phase, training phase and test phase. In training phase, feature vectors are extracted from every speaker’s training data. Then, speaker models are constructed from each speaker’s feature vectors. In test phase, feature vectors are extracted from the test utterance. Then, the speaker model which has the maximum likelihood with the given feature vectors is selected. In general, similar feature vectors are included in different speakers’ training set because of acoustically similar features between speakers, background silence and environment noise. These similar feature vectors cause the overlap of speaker models which contribute to decision errors. As the more speakers are enrolled, the overlapped regions become bigger. Hence it is important to reduce the effect of the overlapped regions. Recently, a feature vector selection method was proposed to mitigate overlap effect. In this system, they classified feature vectors from training data into two categories, non-overlapped and overlapped. Using these separated feature vectors, they constructed non-overlapped and overlapped speaker models for each speaker respectively. In test phase, they only used feature vectors from test utterance which have the maximum likelihood with non-overlapped speaker models. By using this method the system can use only robust feature vectors and it can have better accuracy when speaker models are overlapped. However, a drawback of the previous method is that they didn’t consider the source causing the overlap. If there are more overlapped feature vectors than non-overlapped ones and most of them are caused by acoustic similarity between speakers, the system accuracy will be lowered than conventional method. In this paper...