Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (8): 123-130.DOI: 10.3778/j.issn.1002-8331.1611-0388

Previous Articles     Next Articles

Initial and final segmentation in cleft palate speech based on acoustic characteristics

WANG Xiyue1, HUANG Yipeng1, QIAN Jiahui1, HE Ling1, HUANG Hua1, YIN Heng2   

  1. 1.College of Electrical and Information, Sichuan University, Chengdu 610065, China
    2.West China Hospital of Stomatology, Sichuan University, Chengdu 610041, China
  • Online:2018-04-15 Published:2018-05-02

基于声学特征的腭裂语音声韵母切分

王熙月1,黄毅鹏1,钱佳慧1,何  凌1,黄  华1,尹  恒2   

  1. 1.四川大学 电气信息学院,成都 610065
    2.四川大学 华西口腔医院,成都 610041

Abstract: This paper presents an initial/final segmentation algorithm in cleft palate speech. Through subjective test and objective F test and t test, it is proven that there are significant differences between cleft palate speech and normal speech. Two types of syllables are defined firstly:Class I syllable whose initial has the characteristics of voiceless phoneme, and Class II syllable whose initial has the characteristics of voiced phoneme. These two types of syllables are classified based on hierarchical fuzzy clustering model. Then for the class I syllable, a similar-to-voiced-sound weighting function and similar-to-unvoiced-sound probability function are defined, in order to achieve the roughly initial/final segmentation. The accurate location of initial/final boundary in class I syllable is achieved, through calculating the first-order difference of autocorrelation function’s peak number. For the class II syllable, based on the waveform difference between initial and final, the energy’s jumping point of short-autocorrelation function is found in order to achieve the initial/final segmentation. The experiment results show that the proposed algorithm achieves a high segmentation accuracy rate:Segmentation accuracy for class I syllables is 90.72%, and segmentation accuracy for class II syllables is 92.90%.

Key words: cleft palate speech, initials and finals segmentation, hierarchical clustering, short-time autocorrelation, similar-to-voiced-sounds weighting function, similar-to-unvoiced-sound probability function, short-time magnitude function

摘要: 设计了一种腭裂语音的声韵母切分算法。通过主观的波形测试和客观的F检验及t检验,证明了腭裂语音与正常语音具有显著性差异。定义声母具有清音音素特性的音节为I类音节,声母具有浊音音素特性的音节为II类音节。首先基于层次聚类模型自动判别I类、II类音节,然后定义类浊音权重函数和类清音概率函数,实现I类音节的声韵母一级切分,再通过短时自相关函数峰值个数的一阶微分实现I类音节声韵母的二级切分。基于声韵母波形差异性,检测短时自相关函数的能量跳变点,实现II类音节的声韵母切分。通过大样本实验,结果表明提出的腭裂语音声韵母自动判别算法具有较高的正确率,I类音节的正确率达到90.72%,II类音节的正确率为92.90%。

关键词: 腭裂语音, 声韵母切分, 层次聚类, 短时自相关函数, 类浊音权重函数, 类清音概率函数, 短时幅度函数