Satoru Fukayama

Melody Harmonization With Interpolated Probabilistic Models

Abstract Most melody harmonization systems use the generative hidden Markov model (HMM), which model the relation between the hidden chords and the observed melody. Relations to other variables, such as the tonality or the metric structure, are handled by training multiple HMMs or are ignored. In this paper, we propose a discriminative means of combining multiple probabilistic models of various musical variables by means of model interpolation. We evaluate our models in terms of their cross-entropy and their performance in harmonization experiments. The proposed model offered higher chord root accuracy than the reference musicological rule-based harmonizer by up to 5% absolute.

Statistical Approach to Automatic Expressive Rendition of Polyphonic Piano Music

In this chapter, we discuss how to render expressive polyphonic piano music through a statistical approach. Generating polyphonic expression is an important element in achieving automatic expressive piano performance since the piano is a polyphonic instrument. We will start by discussing the features of polyphonic piano expression and present a method for modeling it based on an approximation involving melodies and harmonies. An experimental evaluation indicates that performances generated with the proposed method achieved polyphonic expression and created an impression of expressiveness. In addition, performances generated with models trained on different performances were perceptually distinguishable by human listeners. Finally, we introduce an automatic expressive piano system called Polyhymnia that won the first place in the autonomous section of Performance Rendering Contest for Computer Systems (RenCon) 2010.

Adaptive aggregation of regression models for music emotion recognition

We present a method for music emotion recognition which adaptively aggregates regression models. Music emotion recognition is a task to estimate how music affects the emotion of a listener. The approach works by mapping acoustic features into space that represents emotions. Previous research has centered on finding effective acoustic features, or applying a multi-stage regression to aggregate the results obtained by using different acoustic features. However, after training regression models, the aggregation happens in a fixed way and cannot be adapted to acoustic signals with different musical properties. We indicate that the most effective feature in estimation differs depending on what kind of emotion we are estimating, and propose a method that adapts the importance of each feature in estimating the emotion. We exploit the variance obtained with Gaussian process regression to measure the confidence of the estimated result from each regression model. The formula for aggregating results from different r...