Improving Generalization for Classification-Based Polyphonic Piano Transcription

In this paper, we present methods to improve the generalization capabilities of a classification-based approach to polyphonic piano transcription. Support vector machines trained on spectral features are used to classify frame-level note instances, and the independent classifications are temporally constrained via hidden Markov model post-processing. Semi-supervised learning and multiconditioning are investigated, and transcription results are reported for a compiled set of piano recordings. A reduction in the frame-level transcription error score of 10% was achieved by combining multiconditioning and semi-supervised classification.