Supervised and unsupervised sequence modelling for DRUM transcription

We discuss in this paper two post-processings for drum transcription systems, which aim to model typical properties of drum sequences. Both methods operate on a symbolic representation of the sequence, which is obtained by quantizing the onsets of drum strokes on an optimal tatum grid, and by fusing the posterior probabilities produced by the drum transcription system. The first proposed method is a generalization of the N-gram model. We discuss several training and recognition strategies (style-dependent models, local models) in order to maximize the reliability and the specificity of the trained models. Alternatively, we introduce a novel unsupervised algorithm based on a complexity criterion, which finds the most regular and wellstructured sequence compatible with the acoustic scores produced by the transcription system. Both approaches are evaluated on a subset of the ENST-drums corpus, and yield performance improvements.