Indexing and Retrieval of Scores by Humming based on Extracted Features

Resumo

In order to be able to conduct searches over large collections of music scores with queries provided in audio format, this article considers recent literature in the field and proposes an implementation to extract specific features from music pieces. Afterwards, we index those features using modern Lempel-Ziv (LZ)-based data structures. These data structures take advantage of the intrinsic repetitiveness within music to reduce space consumption and, at the same time, to index the information optimizing the search time per query. Furthermore, taking advantage of this property-based representation framework, which does not depend on the way the music is portrayed, we enable the possibility to perform melodic searches by simply providing a query audio. This research branch is known as “query by humming” and has commonly been applied to audio sources. A preliminary study for its application in other forms of music representation is presented in this research.