论文部分内容阅读
提出了一种将人耳听觉响度特性应用于噪声下语音识别的前端特征提取方法。本文使用分层遗传算法设计响度加权滤波器,对频谱进行听觉特性加权。将本方法应用于TIMIT数据包的E-SET在NoiseX92的各种噪声条件下的识别实验。实验结果表明,在各种噪声的不同信噪比下,对LPCC和MFCC特征,采用响度加权平均识别率分别有7%和11%的提高,证明本方法是有效的。
A new method of extracting the front-end features of human ear auditory and loudness characteristics is proposed. In this paper, a hierarchical genetic algorithm is used to design a loudness-weighted filter to weight the auditory characteristics of the spectrum. The method is applied to the TIMIT data packet E-SET in NoiseX92 under various noise identification experiments. Experimental results show that the proposed method is effective for LPCC and MFCC features using the weighted average recognition rate of loudness of 7% and 11% respectively under different SNRs of different noises.