Anti-noise speech recognition system based on improved MFCC features and wavelet kernel SVM

作者：Bai Jing^*; Xue Peiyun; Zhang Xueying; Yang Lihong

来源：Advances in Information Sciences and Service Sciences, 2012, 4(23): 599-607.

DOI：10.4156/AISS.vol4.issue23.74

• Anti noise
• Feature parameters
• Gaussian kernel functions
• Generalization ability
• Improved MFCC
• Kernel function
• Low SNR
• Me-xico
• Noisy environment
• Parameters extraction
• Recognition rates
• Speech recognition systems
• SVM
• Wavelet function
• Wavelet kernel
• Wavelet kernel functions

To improve the generalization ability of the support vector machine(SVM) model and solve the problem that recognition rates of the speech recognition system become worse in the noisy environment, in this paper, proposed an improved MFCC parameters extraction method, the new feature MFDWC extracted can accord with human auditory characteristics better and is more suitable under lower SNRs; introduced Mexico wavelet function into the SVM, proved that the Mexico wavelet function can be used as the SVM kernel function, and applied it to Aurora2 speech recognition systems; finally, compared speech recognition rates of SVM based on Mexico wavelet kernel function with those of Gaussian kernel function using traditional MFCC and MFDWC feature parameters respectively. The experimental results show that the SVM speech recognition results of MFDWC- Mexico model are the best in four models (MFDWC-Gaussian, MFDWC- Mexico, MFCC-Gaussian and MFCC- Mexico) under different SNR and different speaker crowds, especially in low SNR, MFDWC model can still maintain a certain amount of recognition rate, further demonstrates that the speech recognition method proposed in this paper not only makes the SVM have good generalization ability, but also have better robustness.

• 出版日期2012
• 单位太原理工大学