Strategies for matching hummed queries

Author: Fábio Goródscy

Advisor: Marcelo Queiroz

Query-by-humming is a common topic in music information retrieval. In the query-by-humming task a hummed record, representing imprecisely a target melody, is given to an application which is supposed to retrieve the target melody from a dataset. Any algorithm addressing this task has to handle the deviations between query and target melodies in both time and frequency domains.

This work reviews standard techniques presented in the academic literature and in commercial applications. Algorithms presented in the conference of the international society for music information retrieval (ISMIR) are reviewed, as well as the commercial application Soundhound, which are explored and tested. This work compares the performance of several strategies for query-by-humming within a unified query dataset.

The main goal of this work is to undercover the difficulties in measuring similarity in this context, by comparing the performance of a commercial tool and also several alternative strategies.