A Study On Speaker Change Detection

With the development of IT technology, rich access to varies type of audio document and thegrowth of data volume, it is more and more difficult to manage the document. In recent years, audiosegmentation and clustering technology have been researched to deal multimedia speech document,the most difficult one is meeting voice. Segmentation and clustering based on speakers is todistinguish voices of speakers and segment speeches into many segmentations in which everyonecontains only one speaker, then mark the same speaker and reset with speaker clustering aftersegmentation. Segmentation is completed primarily with Speaker change detection (SCD), which isto find change time between two different speakers.The segmentation and clustering proposed in the paper consist of three aspects: featureextraction, speaker segmentation and the speaker similarity detection. They are introduced in detailand advantages and disadvantages of different methods are listed with experiments. contents are asfollows:1) Speech feature extraction. LPCC and MFCC are applied for speaker characteristic parameters,through the experiment finds MFCC performance better than LPCC.2) Speaker segmentation. The detection method based on mixed speaker change with credibilitytrend and improved BIC is to utilize credibility trend to solve cumulative error caused by dataaccumulated and to utilize BIC to solve error caused by improper credibility parameterspercentage. The experimental results show that the hybrid algorithm increase by10%and5.8%respectively than their individual use.3) Speaker similarity detection. The paper proposed that gender recognition based on pitch periodand formant, speaker similarity detection and clustering based on GMM model. The proposal isverified to apply to occasions with a small number of people, such as telephone dialog andsmall meeting.