基于改进噪声估计的谱减法应用于说话人识别

Speaker Recognition Using Spectral Subtraction Method Based on Improved Noise Estimation

摘要:

针对语音信号中存在加性噪声使MFCC的鲁棒性和识别系统的性能下降的问题,基本谱减法的引入在增强MFCC抗噪性上取得的效果有限,为了使MFCC具有更好的抗噪性,提出了一种改进算法,在谱减法的基础上引入谱熵的思想,利用谱熵值的分布逐帧进行噪声估计,可更精确地谱减去噪；实验结果表明,当语音中含有加性噪声时,与基本谱减法相比,改进谱减法的说话人识别系统抗噪性与鲁棒性更好。

Abstract:

Aiming at the problem that additive noise in speech signal makes the performance of speaker recognition system degradate when using MFCC. The introduction of traditional spectral subtraction achieved some effect on enhancing noise immunity of MFCC,but the improvement is limited. To get a better result,a novel algorithm of spectral subtract is proposed in this paper. The concept of spectral entropy is introduced based on the spectral subtraction,the noise of each flame is estimated more accurately according to its spectral entropy and subtracted to get better denoising effect. Experimental results show that when there is additive noise in the test speech,compared with traditional spectral subtraction,the speaker recognition system of novel algorithm has better noise immunity and robustness.