Semi-supervised Learning of Bottleneck Feature for Music Genre Classification

摘要：

　　A good representation of the audio is important for music genre classification.Deep neural networks(DNN)enable a better approach to learn the representation of audio.The representation learned from DNN,which is known as bottleneck feature,is widely used for speech and audio related application.However,in general,it needs a large amount of transcribed data to learn an effective bottleneck feature extractor.While,in reality,the amount of transcribed data is often limited.In this paper,we investigate a semi-supervised learning to train the bottleneck feature for music data.Then,the bottleneck feature is used for music genre classification.Since the target dataset contains few data,which cannot be used train a reliable bottleneck DNN,we train the DNN bottleneck extractor on a large out-of-domain un-transcribed dataset in semi-supervised way.Experimental results show that with the learned bottleneck feature,the proposed system can perform better than the state-of-the-art best methods on GTZAN dataset.

关键词： Bottleneck DNN Semi-supervised Multilingual Cross-lingual

作者: Jia Dai Wenju Liu Hao Zheng Wei Xue Chongjia Ni

作者单位: NLPR,Institute of Automation,Chinese Academy of Sciences,Beijing,China;University of Chinese Academy School of Mathematic and Quantitative Economics,Shandong University of Finance and Economics,Shandon

会议类型: 国际会议

会议名称: 第七届全国模式识别学术会议(The 7th Chinese Conference on Pattern Recognition,CCPR2016)

会议地点: 成都

会议语种:英文

页码: 552-562

在线出版日期: 2016-11-03（万方平台首次上网日期，不代表论文的发表时间）