基于多特征后期融合的声学场景分类

Acoustic scene classification based on multi-feature post-fusion

为提高声学场景分类准确率,综合考虑声学事件本身特征对于场景表征的影响以及单模型训练存在的分类误差问题,提出一种基于多特征后期融合的声学场景分类方法。在线性预测倒谱系数的基础上提出声学事件状态似然,结合深度散射谱以及谱质心幅度倒谱系数共同作为特征输入,在残差网络分类器进行并行训练;在分类结果处理阶段,采用平均叠加的整体策略在随机森林上进行集成训练,预测声学场景类别。研究结果表明,所提方法能够有效利用功能互补声学特征对声学场景进行分类,提高分类精度以及泛化性能。

To improve the accuracy of acoustic scene classification(ASC),an ASC method based on multi-feature fusion was proposed considering the influence of acoustic event characteristics on scene representation and the classification error of single model training.Acoustic event state likelihood(AESL)was proposed based on the linear prediction cepstrum coefficient.Combined the depth scattering spectrum(DSS)with the spectral centroid magnitude coefficients(SCMC)as the feature input,the parallel training was carried out in the residual network classifier.In the classification result processing stage,the integrated training on random forest was carried out using the overall strategy of average superposition to predict the acoustic scene category.The study results show that the proposed method can effectively classify acoustic scenes using functional complementary acoustic features and improve classification accuracy and generalization performance.