Cover Song Recognition Using Machine Learning Techniques
Autores
- Andree Silva-Reyes
- Fabiola M. Martínez Licona
- Alma E. Martínez-Licona
Abstract
The task of recognizing a song as a cover version of another is relatively easy for the human being, when the song is known. However, making a machine to do this job is complex because of the number of variables involved in the development of a cover; these include variations in tempo, instrumentation, gender, and duration with respect to the original version. A methodology that aims to identify covers from the application and analysis of machine learning techniques, sparse codification, signal processing and second order statistics, in order to obtain the best configuration, is proposed. Acoustic features such as pitches and timbres, as well as beat information of the cover songs were obtained from the Million Song, a metadata database oriented to music information retrieval. Along the experimentation it was able to try different analysis configurations on the metadata and to appreciate the effects on the comparisons between original and cover versions. According to the results, a system that integrates a frequency processing on the pitches with beat alignment, a sparse codification and a clustering technique was obtained with correct cover identification similar to the state of the art results. It was also possible to get information about learning techniques combinations with different metrics that allows future experiments to improve the results.
Referencia: Andree Silva Reyes, Fabiola M. Martinez-Licona y Alma E. Martinez-Licona (2018). Cover Song Recognition Using Machine Learning Techniques. Research un Computing Science, 147(4):9-21.
Versión completa: Click aquí
Photo by Oleg Magni from Pexels
