Using the audio fingerptint method to create music recommendation systems

Karavashkin L. A., Shkrigunov A. E., Molodyakov S. A., Medvedev B. M.

To create music recommendation systems, two approaches are mainly used: based on metadata and based on audio analysis. We present a joint application of these approaches. A distinctive feature of our recommendation system is the use of audio fingerprints. An audio fingerprint is built on the basis of a spectral analysis of the signal with a preliminary separation of the musical fragment from the voice using a neural network and subsequent determination of the maximum spectral components. The article describes the algorithm for calculating the coefficient of similarity of audio recordings. We have presented a recommendation system that uses a graph database and third-party metadata sources. To build recommendations, similarity relationships obtained after analyzing audio prints and similarity relationships obtained from metadata are used. In conclusion, we present the results of experiments.

Keywords:recommendation system, audio analysis, audio fingerprint, graph database.