Study Of Deep Learning-based Music Automatic Tagging Methods

With the rapid development of the modern information technology,multimedia information resources are increasing dramatically.Music,as one of the most important multimedia resources,has widespread customers and high-demanded market.To make the quality of service in music retrieval and music recommendation meet the increasing user requirement,deep exploration and precise positioning of the music content is of great importance.Music tag is the of music attributes.Music automatic tagging can not only save much manpower and time,but also do good to the standardization and unification of music tags.Therefore,it plays a crucial role in music retrieval and music recommendation.The purpose of this thesis is to study the deep learning-based music automatic tagging methods.The main work is as follows:First,we analyze the music automatic tagging task and study the common-used music tagging methods.We do the deep exploration of MagnaTagATune dataset and analyze the music tags from different dimensions.According to the demand of the music automatic tagging task and the characteristics of the data,we do the tag classification and tag fusion.Mel-spectrogram is extracted from the original music clips as the input signal of the proposed music automatic tagging method.Then,we propose a new convolutional recurrent neural network used for music automatic tagging which combines the convolutional neural network and recurrent neural network parallelly.It uses the convolutional neural network to extract deep features of the mel-spectrogram and uses the recurrent neural network to do the temporal summarization of the mel-spectrogram sequence.With the combination of the outputs of these two neural networks,we can get the prediction probability of each music tag.Finally,during the training of the proposed convolutional recurrent neural network,we set the learning rate dynamically and use a loss function with penalty factor,which enhances the effect of the model training.Receiver operating characteristics curve and the area under it are selected as the main evaluation indicators of the proposed method.Student's t-test between the proposed music automatic tagging model and other models is performed to prove the uniqueness of the proposed method.As the experiment shows,the accuracy of the proposed convolutional recurrent neural network is 95.4%and the area under the receiver operating characteristics curve is 0.903,which outperforms all other existing music automatic tagging methods.