论文部分内容阅读
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 limitations of Mel Frequency Cepstral Coefficients (MFCCs) information in distinguishing audio signals, a two-level classification strategy based on different feature extraction techniques is proposed to classify nine kinds of audio signals of intelligent health monitoring system. The first level of classification uses the MFCCs and their rate of change (ΔMFCCs) as inputs to a Hidden Markov Model (HMM). At the second level, the first-order difference mean and standard deviation of the power spectral density in different frequency bands are taken as the classification features. The experimental results show that the first-order difference of power spectral density contains important classification information which is not contained in MFCCs. This method makes the average classification accuracy of real-time health monitoring system as high as 97.37%, with good robustness and classification accuracy.