Evaluation of a melody transcription system

This paper describes an evaluation of MT, the music transcription front end of the New Zealand Digital Library. The system transcribes sung melodies into music notation, thus enabling users to search collections of musical scores by singing short phrases into a microphone. MT accepts sung input ranging from F2 (87 Hz) to G5 (784 Hz), with note boundaries determined using signal amplitude. Evaluation was carried out by transcribing 100 recorded melodies $approximately 5000 notes. Transcription error was 11.4%, with incorrect note boundaries accounting for virtually all errors. Because of this finding, parameters for the transcription system are set to allow users to clearly separate notes by using “t” at the beginning of each sung syllable