A Dynamic Modelling Approach to Music Recognition
Simon Dixon · 1996
This paper focusses on some of the more difficult issues involved in creating an automatic transcription system. The initial stages of the project follow traditional approaches based on Fourier analysis, but as these methods are not sufficiently robust to process arbitrary musical data correctly, they are augmented by models of auditory perception derived from auditory scene analysis and dynamic models of the sources. We argue that by using dynamic modelling, it is possible to solve many of the constituent problems of automatic transcription. Keywords: automatic transcription, music recognition 1 Introduction Most approaches to machine transcription suffer from brittleness --- they have a steep degradation in performance under non-ideal conditions, such as the minor variations and imperfections in tempo, rhythm and frequency which commonly occur in musical performances. This problem has been noticed in computer implementations of both frequency tracking and beat tracking algorithms,...