一种新的基于分类的音频流分割方法

A Novel Classification-Based Audio Segmentation Algorithm

很多传统的音频流分割方法都是基于小尺度音频分类的,它们普遍存在虚假分割点过多的缺点,严重影响了实际应用的效果.我们的研究表明,大尺度音频片段的分类正确率明显高于小尺度音频片段的分类正确率.基于这个事实和减少虚假分割点的目的,我们提出了一种新的基于分类的音频流分割方法.首先,采用基于大尺度分类的分割方法对音频流进行粗分割,然后采用基于小尺度分类的细分割步骤在边界区域中进一步精确定位分割点.理论分析和实验结果均表明,当处理类别变换频率较低的音频流时,这种分割方法在保持真实分割点检测率的同时能够大幅降低虚假分割率.

Content-based audio segmentation plays an important role in multimedia applications. Many conventional segmentation algorithms are based on small-scale classification and always result in a high false alarm rate. Our experimental results show that large-scale audio can be more easily classified than small ones, and this trend is irrespective of classifiers. According to this fact,we present a novel framework for audio segmentation to reduce the false seg- mentations. First,a rough segmentation step based on large-scale classification is taken to ensure the integrality of the content of segments. Then a subtle segmentation step based on small-scale classification is taken to further locate the segmentation points from the boundary areas computed by the rough segmentation step. Both theoretical analysis and ex- perimental results show that nearly 3/4 false segmentation points can be reduced comparing to the conventional audio segmentation method based on small-scale audio classification, while preserving a low missing rate, when infrequently type-changed audio streams are dealt. So it can be concluded that it is very suitable for the real tasks such as music broadcast segmentation or music video analysis.