Music Detection Using Deep Learning with Tensorflow

Abstract

Music is an expression through collection of harmonic frequencies whose medium is sound. Group of these frequencies will consist of various elements that create music or non music expression. The main objective of the work carried out is to detect the presence of music in a given audio file using the concept of transfer learning. The literature proves that music detection in an audio file can be done by extracting handcrafted audio features like (ZCR, entropy, AMR, LSTER) and train by using classifiers like SVM, Random forest. The advances in machine learning and deep learning architectures have opened the new path for music detection. End to end classification system performs feature extraction and classification jointly this process may lead to extract new unknown feature and contribute to improve the overall accuracy of the system, however to train the CNN networks from scratch we need huge dataset and its time consuming, hence the need of transfer learning ascends. We have used a tensor flow VGGish model released by google as feature extractor which is trained on Audioset data from YouTube videos and finally trained LSTM (Long short term memory) network, a special kind of RNN for classification.