电子学报 ›› 2006, Vol. 34 ›› Issue (10): 1917-1919.

• 论文 • 上一篇    下一篇

汉语普通话的管辖音系学特征及提取方法

刘娇蛟, 贺前华, 韦 岗   

  1. 华南理工大学电子与信息学院,广东广州 510641
  • 收稿日期:2005-11-10 修回日期:2006-03-16 出版日期:2006-10-25 发布日期:2006-10-25

Mandarin GP Features and Extraction Method

LIU Jiao-jiao, HE Qian-hua, WEI Gang   

  1. School of Electronic and Information,South China University of Technology,Guangzhou,Guangdong 510640,China
  • Received:2005-11-10 Revised:2006-03-16 Online:2006-10-25 Published:2006-10-25

摘要: 语音识别中多采用音素作为识别单元,因其数量较多,对神经网络训练复杂度的要求高,在多语言语音识别中需要针对不同语言分别建立识别模块.然而,管辖音系学提出了适用于多语言的语音学特征.本文根据英语和汉语发音的相似性,确定汉语普通话声韵母的GP特征表示形式,并应用神经网络实现特征提取.实验表明,GP特征同样可作为汉语语音的语音学特征.

关键词: GP特征, 特征提取, 普通话, 多语种识别

Abstract: Phonemes are often used in speech recognitions.The amount of phonemes makes the RNN's training complicated.Moreover,different modules need to be built up for different languages in multi-language speech recognition.Government Phonology (GP) provides new primes,which can be used to describe the speech features of different languages.According to the similarity of the pronunciation of English and Mandarin,we introduced GP expressions of Mandarin phonemes and used RNN to extract mandarin GP features.As the experiment results show,GP primes can also be used in describing Mandarin speech features.

Key words: GP features, feature extraction, mandarin, multi-speech recognition

中图分类号: