Chinese Four-tone Receognition of Chinese Speech

• Main Authors: Min-Nan Shiau, 蕭敏男
• Other Authors: Ching-Tang Hsieh
• Format: Others
• Language: zh-TW
• Published: 1993
• Online Access: http://ndltd.ncl.edu.tw/handle/72407897714335253313

碩士 === 淡江大學 === 資訊工程研究所 === 81 === The subject in this paper is to extract the parameter of Chinese tone by two methods for recognizing the Chinese tone variation of continuous speech. The pitch pattern of Chinese isolated words are different from others; for continuous speech, the pitch patterns are not all the same because of the different locative in their article,even they have the same tone.In proceeding this experiment,we segment in the speech signal into Voiced/Unvoiced .The parameter of segmentation is derived from the mean value of spectral envelope band of the first formant, which linear log spectrum is obtained by unbiased estimation.Then, the extraction of pitch is carried out by the cepstrum method in voiced region,and smoothes it by using the median smoothing with 3,5,7 points in cascade.It is necessary to decide the beginning and the end point of the pitch for the influence of adjoining syllable in continuous speech. In this paper we use two method : (1) Slope method (2) Discrimination function to extract the parameters of four-tone. (1) Slope method: including 2-segment slopes(crude structure) recognition. (2) Discrimination function method: using the 2-segment slopes analysis of slope method, which the front slope as X-axis and the latter slope as Y-axis to obtain the distrbution on the Cartesian coordinate and fine discrimination function of four-tone. The speech data of this paper is obtained from 25 sentences (including 196 words which consist of 52 1-tone words,57 2-tone words,29 3-tone words,56 4-tone words and 2 chin-tone words) uttered by 6 men and 4 women speakers.Two groups' speakers are selected to set up the function and test the function, each group consist of 3 men and 2 women. The experiment result identify that discrimination method has the better recognition rate than the slope method.