就算法的执行速度加速为讨论重点,对Mel刻度倒频谱参数(MFCC,Mel-frequency cepstral coefficient)特征撷取的加速方法的各改进步骤详细介绍,其中包涵了快速傅里叶变换、对数运算和在整数运算中维持精准度的方法等主题,并且希望在加速的同时,能顾及到整个系统最后的辨识能力。改进算法比传统MFCC算法的速度提升了约20倍。

Improved Extraction Algorithm for MFCC FeatureYE Qing-yun,JIANG Jia(School of Information Engineering,Wuhan University of Technology,Wuhan 430070,China)This paper discussed the acceleration of execution,detailed every improved step on MFCC(MFCC,Mel-frequency cepstral coefficient) in order to accelerate Feature Extraction,such as FFT,Logarithm and maintaining precision in the integer operation.During acceleration,the final recognition rates of the overall system were also need to be considered.The proposed algorithm could effectively accelerate 20 times the recognition as the traditional ones.MFCC;feature extraction;FFT;logarithm