为连续语音识别用的单词音节神经网络建模的研究

本文主要研究连续语音中单词音节的神经网络建模问题.采用了一种富有特色的特征提取方法,并依据高维空间点覆盖理论,对实际连续数字语音的各不同数字音节,以人工切自连续数字语音中的2640个单字音节,构建连续语音中各不同数字音节的特征空间覆盖区,并使用7308个自连续数字语音中切分出的单字音节,利用仿生模式识别原理,进行了建模正确性验证.验证结果正确率达到97%以上,对同样数量的少量建模样本,识别率优于SVM方法.

Single Figure Syllable Modeling Based on Neural Network for Continuous Speech Recognition

The single figure syllable modeling based on neural network for continuous speech recognition is discussed.A new feature extraction method is used which mainly includes compressing single figure frames according to a certain inter-frame angle,extracting representative information comparing to standard single figure of fixed length.2640 single figure syllables made from continuous speech are used to construct each kind of high dimensional space covering area.By biomimetic pattern recognition theory 7308 single figure syllables made from continuous speech are used to confirm this model in CASSANN-II neural computer and get a quite good result.Experiments show the recognition rate is higher than SVM when the training samples are small.