基于频率段的语音识别算法设计与实现

摘要

Linear prediction cepstral coefficients (LPCC) could well reflect the person' s vocal characteristics, while the mel-frequency cepstral coefficients (MFCC) can be a good simulation of the human ear' s auditory effect. Aiming at different calculation accuracy of mel-frequency cepstral coefficients (MFCC) feature coefficients for speech recognition in different frequency signals and the shortcomings of LPCC for human auditory system. MFCC coefficients and IMFCC coefficients as speech characteristic coefficients of different frequency bands, and combined with LPCC, balanced the distribution of filters are taken, and the entire range of frequency bands are completely covered. The MFCC coefficients and LPCC coefficients are combined as the speech recognition feature extraction parameters.The experimental result proves that the efficiency and the recognition rate of improved algorithm both increases compared to the classical algorithm.%线性预测倒谱参数(LPCC)能很好的体现人的声道特性,而梅尔倒谱参数(MFCC)能很好的模拟人耳的听觉效应.针对MFCC在不同频率段的识别精度不一致和LPCC不能准确模拟人的听觉系统问题,将MFCC参数和IMFCC参数分别作为语音不同频率段的特征参数,结合线性预测参数(LPCC),均衡滤波器的分布,完整覆盖到整个频率段范围,将梅尔倒谱参数和线性预测参数结合起来作为语音识别的特征提取参数.实验结果表明,改进之后的算法从效率上和识别率上都有不同程度的提高.