Studies on Feature Extraction and Discrimination of Spoken Languages.

本研究试图发展一种基于语音基频（Fo）轮廓的口语分类方法。语音数据取自光盘多语言语音数据库。6种语言的语音数据选自亚洲和欧洲语言。首先，用一组多边形直线逼近Fo轮廓，使直线与Fo值的均方误差最小；采用动态规划方法确定了线段的最佳边界。计算每条线路的启动频率、坡度和持续时间。其次，利用主成分分析和判别分析对Fo模式得到的参数进行了分析。采用上述参数的均值和标准差进行分析。结果表明，基于这些参数可以对6种语言进行分类。

This study has attempted to develop a method of spoken language classification based on fundamental frequency (Fo) contours of speech. Speech data ware taken from a CD-ROM Multi-language Speech Database. The speech data of 6 languages were chosen from Asian and European languages. First, Fo contour was approximated by a set of polygonal lines so that the mean square error between the lines and Fo values was minimized ; the optimum boudaries of the lines were determined using a dynamic programming procedure. The starting frequency, slope and duration of each line were calculated. Secondly, parameters derived from Fo pattern were analyzed by using principal component analysis and discriminant analyis. Parameters including mean values and standard deviations of the above parameters were used for the analysis. Results show that the 6 languages can be classified based on these parameters.