This paper describes the prototype of a system for recognition of hummed tunes. The prototype takes as input an unknown hummed tune, turns it into segments representing the distinct notes of the tune, detects the pitch of the segments and turns them into a simple note-like symbolic representation to be matched against a reference database of known tunes. The system returns a list of tunes sorted according to similarity to the hummed tune. Each entry in the list contains information like title, composer, author, and a link to a midi-file enabling the user to ensure himself that the he is reading the data of the song that he has hummed. Testing with seven persons humming 46 songs in total, has shown a successful recognition rate of 75.6% with a search space of 39925 songs.

Download