Fractal dimension feature for distinguishing between overlapped speech and single-speaker speech

Abstract

This paper proposes to distinguish between overlapped speech and single-speaker speech using fractal dimension feature. It is found that the degree of chaos in single-speaker speech frames is lower than that in overlapped speech frames, which indicates that the fractal dimension can be used as a feature to distinguish overlapped speech from single-speaker speech. We carried out experiments for evaluating the effectiveness of fractal dimension. Experimental results show that combining traditional features with fractal dimension feature achieves the highest discrimination rate of 81.0%.