从声学距离计算看汉语声韵母的特点

ANALYSIS OF THE CHARACTERISTICS OF INITIALS AND FINALS IN MANDARIN FROM THE PERSPECTIVE OF ACOUSTIC DISTANCE

本文使用动态时间归整(DTW)算法计算了汉语普通话声母、韵母的美尔倒谱系数(MFCC)声学距离,使用平均组间联接法和邻接法进行了聚类,并做了二维主成分分析。平均组间联接法聚类的结果表明,相同发音部位或结构近似的音往往能在末端节点聚集,鼻韵尾的人际变异性比较大。主成分分析的结果表明,主成分1主要与声韵母的响度/固有振幅有关,主成分2主要体现频谱能量集中区频率和能量持续时间的加和作用。邻接法聚类的结果表明,汉语声母和韵母之间界限清楚,声母类中又依据是否为响音(m、n、l、r)分为两类。非响音中发音部位和发音方法交替作用主导分类,但塞音和塞擦音、擦音总是截然分开的,塞音和塞擦音中,送气与否的作用相对较小。韵母中齐齿呼、撮口呼韵母与开口呼、合口呼韵母的界限较为清晰。在开口呼和合口呼韵母中,往往又会依据韵腹发音部位的近似程度聚为不同的小类。此外,在汉语声母、韵母的聚类中,邻接法比平均组间联接法效果更好。

In previous studies on the similarity of the initials and finals in Mandarin,researchers have typically calculated the perceptual distances,while the acoustic distances among those phonemes were less studied.In addition,the investigation on the initials is more often seen than that on the finals in terms of both the perceptual distance and the acoustic distance.While most of these studies have analyzed the initials and finals separately,only a limited number of them have compared the initials and finals on the same level,which is regarded to be useful to understand the characteristics of the initials and finals in Mandarin.In this paper,the initials and finals in Mandarin were placed in the context of“X-Y-X”,where X could be one of the six monophthongs(i.e.,/a oɣi u y/),and Y could be any of the the 21 initials and 39 finals in Mandarin.We collected recording data from four speakers(two males and two females)who were asked to read the segmental combinations in a sound-proofed laboratory.All speakers are trained students in phonetics.The“Dynamic Time Warping”algorithm was used to calculate the acoustic distance(based on Mel-frequency cepstral coefficients)of the initials and finals in Mandarin.According to the distance matrix based on acoustic distance data,the Average Linkage(between-group)method and Neighbor Joining method were used to cluster,and the two-dimensional Principal Component Analysis was carried out.The clustering results of the Average Linkage method showed that sounds with the same place of articulation or similar structure often gathered at one terminal node,but the interpersonal variability of nasal ending was relatively large.The results of PCA showed that 1st component was mainly related to the loudness of sounds or the intrinsic amplitudes of sounds;2nd component mainly reflected the additional effect of frequency and duration of energy in the spectral energy concentration region.The results of Neighbor Joining clustering showed that the boundary between the initials and the finals is clear,and the initials are further divided into two categories according to whether or not it is a sonorant(i.e.,m,n,l,r in this study).The place and manner of articulation alternately played a leading role in the classification of non-sonorants.Stops were always completely separated from affricates and the fricatives.For affricates and fricatives,the role of aspirating was relatively unimportant.In the category of finals,the Qichihu finals and the Cuokouhu finals were clustered into one group,while the Kaikouhu finals and the Hekouhe finals were clustered into another group.The Kaikouhu finals and the Hekouhu finals were often divided into different sub-categories according to the main vowels.Last but not least,we showed that the Neighbor Joining method was better at clustering of the Chinese initials and finals than the Average Linkage(between-group)method.