Combining Novel Acoustic Features using SVM to Detect Speaker Changing Points

Abstract

Automatic speaker change point detection segments different speakers from continuous speech according to speaker characteristics. This is often a necessary step before applying speaker verification or identification systems. Among the features to represent a speaker in the speaker change point detection systems acoustic features are commonly used. Commonly used features are Mel Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC). However, the features are affected by speech content, environment, type of recording device, etc. So far, no features have been discovered, which values depend only on the speaker. In this paper four novel feature types proposed in recent major journals and conference papers for speaker verification problem, are applied to the problem of speaker change point detection. The features are also used to form a combination scheme via SVM classifier. The results shows that the proposed scheme improves the performance of speaker changing point detection as compared to the system that uses MFCC features. It was also found that some of the novel features of low dimensionality give comparable speaker change point detection accuracy to the high-dimensional MFCC features.