基于EEMD的异常声音特征提取

Abnormal Sound Feature Extraction Based on EEMD

针对使用梅尔倒谱系数MFCC,LPCC等传统语音特征时异常声音识别率低的问题,结合到异常声音具有高度非平稳、非线性的特点,提出一种基于总体平均经验模态分解的异常声音特征提取方法。首先对声音进行分帧,对每一帧信号提取模态函数。对不同层模态函数提取包括短时能量,能量比,短时平均过零率,MFCC等特征,对信号的特征向量分段取均值作为最终的特征。基于这些特征的特征组合,采用支持向量机作为分类模型对七种异常声音进行识别,并测试了不同信噪比条件下识别的效果,结果表明基于EEMD的特征相比MFCC,LPCC等特征能有效提高识别率。

Aiming at solving the low recognition rate of abnormal sound recognition caused by using MFCC, LPCC as feature, the project proposes a feature extraction method for abnormal sound based on Ensemble Empirical Mode Decomposition （EEMD） combining the high nonlinearity and non-stationary. First the abnormal sounds are segmented into frames and every frame of the sound is decomposed into IMFS, then features including energy, cross rate, energy ratio, and MFCC are extracted for every IMF. Finally the feature vectors are segmented and the means of every segment are computed as the final features. Using these features as input, then the project adopts SVM as classifier to recognize seven kinds of abnormal sounds, and the recognition rate is tested in railway background. Experiment results show that these features can improve the recognition rate comparing with MFCC.