Text-Independent Speaker Identification Utilizing Likelihood Normalization Technique
Konstantin Markov, Seiichi Nakagawa · IEICE Transactions on Information and Systems · 1997
SUMMARY In this paper we describe a method, whichallows the likelihood normalization technique, widely used forspeaker verification, to be implemented in a text-independentspeaker identification system. The essence of this method is toapply likelihood normalization at frame level instead of, as it isusually done, at utterance level. Every frame of the test utter-ance is inputed to all the reference models in parallel. In thisprocedure, for each frame, likelihoods from all the models areavailable, hence they can be normalized at every frame. A specialkind of likelihood normalization, called Weighting Models Rank,is also experimented. We have implemented these techniques inspeaker identification system based on VQ-distortion codebooksor Gaussian Mixture Models. Evaluation results showed thatthe frame level likelihood normalization technique gives higherspeaker identification rates than the standard accumulated like-lihood approach.key words: speaker identification, likelihood normalization,frame level processing