A Learning-Based Quantization: Unsupervised Estimation of the Model Parameters

Abstract

This paper describes a method for organizing onset times performed along a jam-session accompaniment into normalized (quantized) positions in a score so the performance data can be stored in a reusable form. Unlike most previous beat-tracking-related methods that focus on predicting or estimating beat positions, our method deals with the problem of eliminating the onset-time deviations under the condition that the beat positions are given. Our method solves this problem by using hidden Markov models (HMMs) that model onset-time transition and deviation. The HMM parameters are obtained by unsupervised estimation using the Baum-Welch algorithm and held-out interpolation: they can be derived from only the session recording that we wanted to quantize. Experimental results show that our model performs better than the semi-automatic quantization in commercial sequencing software. 1