A Study On Speaker Diarization Based On Multiple Features

Speaker diarization is the task of automatically partitioning an input audio stream into homogeneous segments without any prior knowledge and annotating the audio with the speaker identity. Speaker diarization has many applications such as speaker adaptation,speaker tracking and speaker indexing and so on. It mainly involves two speech processing issues: speaker segmentation and speaker clustering.Now speaker diarization is mainly based on the statistical methods in the absence of prior knowledge ,information of the speaker and the number of the speaker, leading to too little sample information, and ultimately affect the accuracy of the method.Based on the existing research achievement of speaker diarization ,we suggested a speaker diarization system framework based on the methods of multiple features and multilevel aiming at solving the problems caused by the lack of available information .We carries out the following researches on the related algorithms.Firstly,the key technologies of speaker diarization system are described and analyzed forcusing on the algorithm of analysis and extraction of feature,speaker model, speaker segmentation, speaker clustering.Secondly, to take full advantage of the information of the speaker ,a new method is proposed which combing vocal track features and sound source features ,meanwhile we also use the method of feature selection.The purpose of these methods is making full use of limited information and enhancing the performance of the system.Thirdly,we investigate that the choice of different distance measure has an effect on speaker segmentation which reduce the error accumulation caused by segmentation error. We also get more into the way that the research is a new segmentation algorithm based on one-class SVM and experiments show that the proposed algorithm is effective.Fourth, to solve the problem of speaker information deficiency affecting accuracy of segmentation in the conventional speaker diarization system,we develop a re-segmentation and re-clustering scheme using method of viterbi decoding based on EHMM combine with unsupervised segmentation algorithm based on distance measure and a hierachical clustering algorithm using BIC. Experimental results show that the system has good performance...