Music transcription is the process of resolving the musical score from an audio recording. The ability to generate an accurate transcript of a musical performance has numerous practical applications ranging in nature from content-based organization to musicological analysis. Although trained musicians can generally perform transcription within a constrained setting, the process has proven to be quite challenging to automate since the recognition of multiple simultaneous notes is generally obfuscated by the harmonic series interaction that renders music aurally pleasing.

In contrast to model-based approaches that incorporate prior assumptions of harmonic or periodic structure in the acoustic waveform, we present a classification-based framework for automatic music transcription. The proposed system of support vector machine note classifiers temporally constrained via hidden Markov models may be cast as a general transcription framework, trained specifically for a particular instrument, or used to recognize higher-level musical concepts such as melodic sequences. Although the classification structure provides a simple and competitive alternative to model-based systems, perhaps the most important result of this thesis is that no formal acoustical prior knowledge is required in order to perform music transcription.

We report a series of experiments, with corresponding comparisons to alternative approaches, in which the proposed framework is used to transcribe real-world polyphonic music ranging in diversity from ensemble orchestral recordings to popular music tracks. In addition, we describe several methods for extending a limited set of labeled training data, thereby improving the generalization capabilities of the classification system. Finally we relate a demonstrative experiment in which the classification posteriors (i.e. the outputs of the proposed framework) are used as an acoustic feature representation.