正交约束型SincNet可微分前端及在音频分类中的应用  

Differentiable front-end of SincNet with orthogonal constrainsand its application in audio classification

作　　者：刘伟 孙蒙 张玥 张雄伟 Liu Wei;Sun Meng;Zhang Yue;Zhang Xiongwei(Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]中国人民解放军陆军工程大学,江苏南京210007

出　　处：《网络安全与数据治理》2022年第8期92-98,共7页CYBER SECURITY AND DATA GOVERNANCE

基　　金：江苏省优秀青年基金(BK20180080);国家自然科学基金(62071484)。

摘　　要：在音频场景分类任务中,现有端到端模型中特征建模层学习过程存在缺乏约束、学习结果缺乏直观解释以及仅适用于特定的后端分类模型等缺点。因此,以SincNet可微分前端为基础,引入正交约束提高其求解效率,同时提高所得可微分前端的可解释性,并使其能提高多种后端网络的分类性能。提出的这种新型可微分前端命名为正交约束型SincNet(OrthSincNet)。研究发现,OrthSincNet卷积核对应的频谱既具有类似于梅尔滤波器的形态,又可提高分类效果。在UrbanSound8K官方评测数据集上的实验表明,相对于常用的梅尔频谱前端,OrthSincNet改进了6种后端分类网络的性能,分类准确率平均提高了2.2%。In the audio scene classification task,the feature modeling layer learning process in the existing end-to-end models has shortcomings such as lack of constraints,lack of intuitive interpretation of learning results,and only applicable to specific back-end classification models.Based on the differentiable front-end of SincNet,this paper introduces orthogonal constraints to improve the performance of the solution of the overall network,tries to promote the interpretability of the learnt front-end,and adapts it to multiple back-end classifiers.The proposed new front-end is called Orthogonal SincNet(OrthSincNet).Therefore,the convolution kernels of OrthSincNet is expected to not only hold good properties on frequency responses like Mel filters but also to be able to tune the parameters adaptively for specific back-end classifiers.Experimental results on the official test set of UrbanSound8K showed that OrthSincNet improved the conventional Mel filter banks by 2.2%in average on 6 back-end classifiers.

关 键 词：可微分前端 正交卷积 SincNet 端到端分类网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]