Vocalist Identification in Audio Songs Using Convolutional Neural Network

There are millions of songs in the world. Manually, it is a cumbersome task to classify songs according to the vocalist. On the other hand, search algorithms can efficiently recognize a song but these algorithms use pattern matching technique in a database and therefore cannot be generalized over known songs. Music companies often face these types of challenges in automation of song classification. Hence, there is a need for artificial intelligence-enabled method to classify songs based on vocalists which can also be generalized for unknown songs. This paper presents a deep learning architecture to achieve the objective of identifying vocalists in audio songs. For that purpose, a novel bollywood song dataset (BSdataset) is created. The pre-processing of dataset involves the conversion of an audio file into a spectrogram, i.e. visual representation of frequencies of audio signal as it varies with time and then uses these spectrograms as an image to train a convolutional neural network (CNN) for classification of vocalists in an audio song.