Automatic transcription of music

This thesis concerns the transcription of music, where the idea is to pick a musical performance by a microphone or from a musical recording, and convert it into a symbolic representation. According to musical practice, this requires extraction of notes, their pitches, timings, and classification of the instruments used. The respective subproblems, pitch tracking, rhythm detection and the analysis of musical instruments, were studied and are reviewed. Emphasis of this thesis is laid on two points: on literature review and on algorithm development. A literature review was conducted on automatic music transcription and several related areas of interest. An appropriate decomposition of the problem and the selection of an approach is first considered. Then the state-of-the-art of the research is represented and discussed, and promising directions for further work are indicated. An original and the most important part of this thesis concerns the development of algorithms that can be used to detect and observe harmonic sounds in polyphonic signals. A novel number theoretical method is proposed, which is motivated by the spectral properties of a mixture of harmonic sounds, especially in musical signals. The performance of the method is evaluated by applying it in a piano music transcription program, which was implemented and simulated in Matlab environment. In the last chapter, the role and use of internal models and predictions in music transcription are discussed. This part of the work introduces processing principles that utilize high-level knowledge sources, such as instrument models and perception-based dependencies in music.

• music