基于多特征融合的SVM声学场景分类算法研究

针对DCASE2017挑战赛的声场环境数据集，提取梅尔频率倒谱系数（MFCC）、短时能量（SE）、声学事件似然特征（AELF）、静音时间（MT）特征，组成多特征融合矩阵，通过对比多种核函数和寻优算法，最终选取高斯径向基核函数（RK）建立支持向量机（SVM）模型，采用交叉验证（CV）方法进行SVM参数寻优，对15种声学场景进行分类.实验结果表明，杂货店、办公室的分类准确性达到了90%以上，平均分类准确性达到71.11%，远高于挑战赛的基线系统61%的平均分类准确性. 

Support Vector Machine for Acoustic Scene Classification Algorithm Research Based on Multi-Features Fusion

For the sound environment dataset of the DCASE 2017 Challenge, Mel frequency cepstral coefficients (MFCC), short-time energy (SE), acoustic event likelihood features (AELF), and mute time (MT) features were extracted to form a multi-features fusion matrix. Comparing various kernel functions and optimization algorithms, radial basis function kernel (RK) was finally selected to establish the support vector machine (SVM) model, and cross validation (CV) method was utilized to optimize SVM parameters and to classify 15 acoustic scenes. The experimental results show that the classification accuracy of grocery store and office can reach more than 90%, and the average classification accuracy reaches 71.11%, which is much higher than the average classification accuracy of 61% of the baseline system given in the challenge.