Deep Analysis On Labels For Music Auto-Tagging

With the rapid development of music social media,there are more and more online music resources and music types.As an effective way to organize massive music data,music annotation has become an important part of online music services.At the same time,more and more research study music auto-tagging,which is an effective method of enriching music annotations.In recent years,deep neural networks have been successfully applied to music tagging.But there are a large amount of label noises in the real music tagging dataset,which reduces the reliability of networks training and analysis,resulting low music tagging performance.The adverse effect of label noises cannot be directly solved by deep neural network structure optimization.In conclusion,the main works of this article are as follows:1)We analyze a large music tagging dataset.And we investigate the effects of label noises on convolutional neural network training to deepen our understanding of their properties.It is found that the label noises makes the network learn a more fragmented mapping of the input,which leads to low performance of final tagging.2)Define the label vector and verify its feasibility and validity.We define the label vector to find the deep relationship between music and tags based on above analysis.Label vector is applied for music auto-tagging task.3)We propose a music auto-tagging algorithm based on deep analysis on labels,which extracts multi-level features of audio and label vector to improve the robustness of the music tagging network and be effective on the real datasets.The experiments on Million Song Dataset shows that we can get the deep relationship between music and tags from pre-trained network.In addition,the experiments show that the proposed method outperforms previous state-of-the-arts on the Magna Tag ATune dataset,the Mil ion Song Dataset and a real business dataset.