【摘要】 该介绍了中文广播新闻语料识别任务中的自动分段和自动分类算法。提出了3阶段自动分段系统。该方法通过粗分段、精细分段和平滑3个阶段,将音频流分割为易于识别的音频段。在精细分段阶段,文中提出两种算法:动态噪声跟踪分段算法和基于单音素解码的分段算法。仿效说话人鉴别中的方法,文中提出了基于混合高斯模型的分类算法。该算法较好地解决了音频段的多类判决问题。在“新闻联播”测试数据中的实验结果表明,该文提出的自动分段和分类算法性能与手工分段分类性能几乎相当。更多还原

【Abstract】 This paper describes the work on the development of an audio segmentation and classification system applied to a broadcast news task for Chinese language. Three-phase automatic audio segmentation algorithm is provided. Audio stream is cut to audio segments (or sentences) by simply segmentation, fine segmentation and smoothing. Two different fine segmentation algorithms are given. They are dynamic noise tracking segmentation algorithm and segmentation based on mono-phone decoder algorithm respectively. Classifier based on mixture Gaussian model is used to classify audio segment into four groups: noise, music, male and female. The experiments on “Xin Wen Lian Bo” broadcast news show the performance of automatic segmentation and classification is almost equivalent to that of manual segmentation and classification.更多还原