基于Tucker分解的音频分类研究

摘要

In this paper,obtaining robust different properties of audio signal by Tucker decomposition was presented and applied to audio classification via Gaussian Mixture Model (GMM).Firstly,acoustics feature sets,perceptual feature sets and psychoacoustic feature sets were extracted after pre-processing.Next,a 3-order Tucker tensor was constructed by the three sets.Projection matrices and principal components in each mode were obtained via Tucker decomposition.Finally, the feature sets were measured for 300 audio recordings consisting of 3 different classes including music,speech and noise. A class of supervised classifiers was developed based on GMM.The experiments compared the percent correct classification alongside different feature sets.The results demonstrate the features via Tucker decomposition outperform the traditional feature sets in the aforementioned experiment.%提出一种利用Tucker分解获得鲁棒性较强的音频信号不同属性的特征，在高斯混合模型上测试音频信号分类性能的方法。音频信号经过预处理后，提取其不同类型特征集合，包括常规声学特征参数集合、听觉感知特征参数集合、心理声学特征参数集合；然后由三种特征集合构建三阶特征张量，通过Tucker分解得到每一类特征阶投影矩阵并进行主分量分析；最后使用包括音乐、语音、噪声3种类型的300条音频数据测试不同特征集合的分类效果，在此过程中使用了有监督学习的高斯混合模型作为分类器。实验中比较了不同特征集合使用高斯混合模型的分类正确率。实验结果表明，Tucker分解获得的特征集合实现了较好的分类，说明该方法性能优于传统特征集合。