Problem-Solving Strategies in a Music Transcription System

Abstract

Music transcription is a significant problem in machine perception. The system discussed in this paper, MANA. takes as input either the sound of a recorded performance, or data captured using a musical keyboard. It produces conventional musical notation as output. It has successfully handled pieces ranging from 18th-century piano music to improvisations on conga drums, in the Afro-Cuban style. The paper describes the key ideas and techniques found in the temporal analysis component of MANA, the goal of which is to assign a rhythmic value to each note played. Perception seems to result from the interplay of sensory evidence and pre-existing mental structures, data-driven agents pitted against model-driven agents in the formation of hypotheses that seem tolerable to our prejudices and not too distant from the data. To this first order view one must add a second order one, which allows the system to notice patterns among partially elaborated hypotheses, and to use these patterns to alter confidence ratings, according to peer pressure rules. Concerning the use of multiple criteria for hypothesis evaluation and pruning, which necessary in the context of perception systems, it is argued that the use of partial orders over multi-dimensional spaces of natural criteria yield more robust methods that approaches based on combining criteria down to scalars.