采用多特征组合优化的语音特征参数研究

Study of Speech Characteristic Parameters by Optimized Multi-feature Combination

在研究传统语音特征参数线性预测倒谱系数(LPCC)和梅尔频率倒谱系数(MFCC)的基础上,加入基于人耳听觉特性的Bark子波滤波器组所提取的特征参数,来共同组成特征集。同时将基于遗传算法的相关性特征算法将特征集进行优化,分别采用贝叶斯和径向基神经网络算法进行语音识别。实验结果表明本方法与传统的LPCC和MFCC方法相比,平均识别率分别提高了4.66%和3.5%,最佳达到98.1%的识别率。

Based on studying traditional speech characteristic parameters of LPCC （linear prediction cepstral coefficients） and MFCC （Mel Frequency cepstral coefficients）, the characteristic parameters based on human auditory characteristics extracted by Bark Wavelet filter groups is proposed, thus to form the feature set. Correlation algorithm based on genetic algorithm is used to optimize the feature set, then the Bayes and the Radial Basis Function neural network algorithm are respectively adopted for speech recognition. The experimental results show that this method, as compared with LPCC and MFCC method, could raise the average recognition rate by 4.66% and 3.5%, and even could achieve 98.1% recognition rate in best case.