摘要
提出了一种基于Bark子波变换和概率神经网络(PNN)的语音识别模型。利用符合人耳听觉特性的Bark滤波器组进行信号重构并提取语音特征,然后利用训练好的概率神经网络进行识别。通过训练大量语音样本来构成语音识别库,并建立综合识别系统。实验结果表明该方法与传统的LPCC/DTW和MFCC/DWT方法相比,识别率分别提高了14.9%和10.1%,达到了96.9%的识别率。
The paper proposes a speech recognition model based on Bark Wavelet Transform and Probabilistic Neural Network （PNN）.According to the group of filters which is similar to human auditory system,reconstruct the signal to abstract speech features then do the recognition work by trained PNN.By training a large number of speech samples,speech identification database is constructed,and the integrated recognition system is then built.The experiment results show that comparing with traditional ways of LPCC/DTW and MFCC/DTW,this method can increase the recognition rate by 14.9% and 10.1% ,and it can attain recognition rate of 96.9%.
出处
《计算机工程与应用》
CSCD
北大核心
2007年第19期30-31,44,共3页
Computer Engineering and Applications
基金
国家自然科学基金(the National Natural Science Foundation of China under Grant No.60572076)
江苏省高校自然科学研究计划项目(the Natural Science Research Project of Higher Education of Jiangsu Province of China under Grant No.05JKB510113)
关键词
Bark子波
概率神经网络
特征提取
语音识别
Bark wavelet
Probabilistic Neural Network （PNN）
feature abstraction
speech recognition
