1-05 - Connective Fusion: Learning Transformational Joining of Sequences with Application to Melody Creation

Taketo Akama

Keywords: Domain knowledge, Machine learning/Artificial intelligence for music, Applications, Music composition, performance, and production

Abstract: We present Connective Fusion, a music generation scheme by transformational joining of two musical sequences for creative purposes. Given two shorter sequences as inputs, our model transforms each of them such that their concatenation is more coherent to form a longer sequence, while each of the transformed shorter sequences retains meaningful similarity with the corresponding input sequence. In short, our model connects and fuses two contextually unrelated sequences in a coherent way. This transformation can be applied iteratively to gradually fuse the input sequences. The style latent space is simultaneously learned, allowing users to control how the two sequences are merged. Our approach comprises two steps of unsupervised learning: a deep generative model with a latent space is learned, followed by adversarial learning of the transformation function in the latent space. We demonstrate the usefulness of our method through the task of melody creation using a symbolic music dataset.