Guiding Factor Oracle Improvisation With a Probabilistic Model

Here, we compare improvisations generated with a Factor Oracle alone vs. improvisations generated with a Factor Oracle combined with a probabilistic model. The probabilistic model is an optimised interpolation of smoothed sub-models learnt on a corpus.

In the following examples, the probabilistic sub-models used are a bigram on the melody and a model representing the correlations between the melody and the harmony.

The training corpus is Charlie Parker's Omnibook (database available here). It has been divided into 3 sub-corpora : a training corpus to estimate the sub-models probabilities; a validation corpus to optimise the interpolation and smoothing coefficients of the probabilistic model; and a test corpus.

The Factor Oracles are constructed on a single tune from the test corpus.