Boosting audio chord estimation using multiple classifiers
Matevž Pesek, Aleš Leonardis, Matija Marolt · University of Birmingham Research Portal (University of Birmingham) · 2014
The paper addresses the task of automatic audio chord estimation using stacked generalization of multiple classifiers over Hidden Markov model (HMM) estimators. We evaluated two feature types for chord estimation: a new compositional hierarchical model and standard chroma feature vectors. The compositional hierarchical model is presented as an alternative deep learning approach. Both feature types are further modelled with two separate Hidden Markov models (HMMs) in order to estimate chords in music recordings. Further, a binary decision tree and support vector machine are proposed binding the HMM estimations into a new feature vector. The additional stacking of the classifiers provides a classification boost by 17.55% with a binary decision tree and and 21.96% using the support vector machine.