Music Similarity Estimation based on Feature Extraction and Deep Learning

Abstract

The field of music information retrieval (MIR) has developed rapidly in recent years, which provides a lot of convenience for people's entertainment life after combining with big data and deep learning. This paper studies music similarity estimation in the field of music information retrieval. To be specific, based on the deep learning of music feature data, the music similarity model (a classifier) is obtained to judge whether the music similarity reaches more than 50% (i.e., random model). According to the experiment, using the Siamese neural network to train the model with chroma features as input can achieve high accuracy. It will make it more convenient for music recommendation, humming recognition and plagiarism identification. In the future, more researches ought to focus on the real value instead of classifier to judge the similarity or at least the clustering model (multiple thresholds instead of two). Anyway, these results shed light on guiding further exploration focusing on similarity extraction of music in the context of the state-of-art deep learning scenarios.