Music Recognition Algorithms Using Queries by Example

With the appearance of audio databases novel information retrieval methods are required. A natural way of searching in a musical audio database is by humming or whistling the tune of a song as a query, which is the so-called “query by humming”. In this study five different techniques for effective and efficient querying by humming are described and compared. For this purpose, a system that allows searching for songs in a pre-prepared database using various pitch tracking methods was implemented. The system uses the hummed part of a song or recording of live singing as an input signal for which melodic contours are created. These contours are then compared with sample songs in the database. Five different algorithms were implemented to search the pitch of the input signals and tested in order to determine which of them is the most effective. Additionally, research has been made to set out influence of predefined parameters and the singer vocal skills on search results.