MusÆ: an Adversarial Autoencoder for Music

Automatic music modelling and generation is a challenging task. The ability to learn from big data collections of deep generative models makes them well-suited for modelling musical data. Among them, the adversarial autoencoder model stands out for its intrinsic flexibility and seems to be a natural choice for dealing with complex data distributions, such as the one of music. Despite that, in the literature there are no mentions of adversarial autoencoders applied to music. This thesis intends to fill this gap, presenting a novel architecture for symbolic music generation, called MusÆ. The experiments show that MusÆ has a higher reconstruction accuracy than similar models based on standard variational autoencoders. It is also able to create realistic interpolations between two musical sequences, smoothly changing the dynamics of different tracks. Experiments on the learned latent space show that some latent dimensions have a significant correlation with some low-level properties of the songs, allowing us to perform changes to the generated pieces in a principled way. We encourage the reader to judge the quality of results by uploading a selection of the generated songs on YouTube