摘要
针对易混淆词特征差异小、分类决策困难的特点,提出了一种新的语音识别特征.该特征可以根据待识单词的发音特点,通过选用合适的基函数及加权处理,突出混淆单词特征之间的差异性;同时,根据其矢量维数相等的特点,利用静态神经网络分类决策能力强、容错性好的优点进一步提高系统的识别性能.实验结果表明,所用方法比传统的DHMM方法和其他神经网络语音识别方法具有更好的识别效率.
This paper presents a novel feature (Weighted Global Time Frequency feature, i.e WGTF) for confusing word speech recognition, which enhances the difference among different confusing words by selecting proper base fuctions and weighting functions. Meanwhile, the storng discriminative power of artificial neural network has been used as a classifier to further raise the recognition rate. The experiment shows that the proposed method outperforms the standard DHMM and other ANN based method.
出处
《应用科学学报》
CAS
CSCD
1998年第3期320-325,共6页
Journal of Applied Sciences
基金
国家攀登计划认知科学(神经网络)重大关键资助
关键词
易混淆词识别
语音识别
全局时频特征
DHMM
vocal tract model, weighting function, GTF feature, speech recognition, neural network