智能保健监测系统中音频信号的分类算法研究  

Audio Signal Classification Algorithm for a Smart Health-Care Monitoring System

作　　者：李玲俐[1] 

机构地区：[1]广东司法警官职业学院信息管理系,广州510520

出　　处：《重庆师范大学学报（自然科学版）》2012年第4期73-76,共4页Journal of Chongqing Normal University:Natural Science

基　　金：广东省自然科学基金(No.101754539192000000)

摘　　要：针对Mel频率倒谱系数(MFCCs)信息在区分音频信号时的局限性,提出一种基于不同特征提取技术的两级分类策略,对智能保健监测系统的9种音频信号进行分类。分类的第一级采用MFCCs及其变化率(ΔMFCCs)作为隐马尔可夫模型(HMM)的输入。在第二级,将不同频段的功率谱密度的一阶差分均值和标准差作为分类的特征。实验结果表明,功率谱密度的一阶差分包含了MFCCs所不含有的重要分类信息,该方法使得实时保健监测系统的平均分类准确度高达97.37%,具有较好的鲁棒性和分类准确性。Aiming at the deficiency of Mel-frequency cepstral coefficients （MFCC, s） in discriminating acoustic signals, a two level classification strategy based on different feature extraction techniques was proposed to classify nine audio signals in a smart health care monitoring system. In the first level, the MFCCs and its variants （ΔMFCCs） are used as the inputs of the hidden Markov model （HMM） for classification. Then, the mean and standard deviation of the first-order difference of power spectral density over different frequency bands are calculated as features for further classification in the second step. Experi- ment results in real-time health monitoring system reveal that the first-order derivatives of power spectral density contain some important information which is not included in MFCCs. The approach in this paper shows better robustness and high classifica tion accuracy whose average is as high as 97.37%.

关 键 词：音频信号 MEL频率倒谱系数 特征提取 隐马尔可夫模型 分类 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]