Automatic Selection and Concatenation System for Jazz Piano Trio Using Case Data

Abstract

In this paper, we discuss a computational model of an automatic jazz session system that is statistically trainable model using lead sheet and jazz session data, in addition, we provide an implementation as prototype system based on this model. In contrast to most previous jazz session systems that required heuristic rules and the human labeling of training data to estimate musical intention of human players, we suggested a statistically trainable mathematical model of jazz session using stochastic state transition model approximating a musical trajectory model. Based on the model, we developed a jazz session system as a prototype using concatenation of case data from real jazz session recordings to show the validity of our model. This system consists of training phase and concatenating phase. In the training phase, the system learns some parameters to classify the piano, bass, and drums data using non-negative matrix factorization, and calculates the chain probabilities by trigram and co-occurrence probabilities between piano, bass, and drums. In the concatenating phase, the system estimates musical states of bass and drums from piano midi-format input, searches and selects a suitable musical data from case data, and concatenates a musical data matching the key between input piano and bass. As a result of the comparative evaluation experiment using some concatenated midi-format data by above methods, our system was found to generate a jazz piano trio musical data having naturalness and shown validity of our proposing model.