改进的音频混合分割方法

Research on the Improved Hybrid Segmentation Algorithm for Audio

针对基于距离和贝叶斯信息准则的混合分割算法在候选分割点确认时存在过于激进、容易造成分割点丢失的问题,提出一种保守的分割点确认方法,使被否定的候选分割点有多次机会被检验;针对固定的惩罚因子无法兼顾准确率和查全率的问题,提出了基于可检测度的惩罚因子自适应算法,并在一个启发式规则的基础上对基于可检测度的惩罚因子自适应进行扩充,实现了基于可检测度和启发式规则的惩罚因子自适应方法.实验结果表明,文中算法明显优于已有算法,且在性能上得到了很大提升.

DISTBIC is a typical hybrid audio segmentation algorithm, but its validation method for candidate changing points is too radical and may lose the true changing points. To address this problem, we develop a new BIC validation algorithm, which gives the candidate points several opportunities to be validated. For a fixed penalty factor value, it is difficult to set a proper value to achieve both high recall rate and high precision. Based on the detect ability and heuristic rule, an adaptive setting method of the penalty factor is designed to solve this problem. The experimental results indicate that the improved algorithm is superior to the original algorithms.