DeepDIVA automatic synthesizer programmer project

DeepDIVA is a data science project in which we aim to create a deep-learning syntheziser programmer using the vst plugin DIVA. We are creating a tool that addresses the challenge of programming sound synthesizers. This requires a thorough technical understanding as software synthesizers often have many parameters. To tackle this, we built a deep learning model for automatic synthesizer programming that takes a target sound as input, predicts the parameter settings for the synthesizer that cause it to emit as close a sound as possible, and generate a template that you can load in your DIVA synthesizer. In future work, we aim to improve the sound matching performance of our deep neural network architecture. At this point in time the sound matching performance of our best performing architecture is too limited for an application that users can integrate into their workflow, at least not for a synthesizer as complex as the Diva.

We generated and saved datasets for training and validating deep learning models. Here we share the dataset (extracted features and patch targets) we created on which our currently best performing model is trained. For this dataset, 11 parameters of the DIVA synthesizers were allowed to vary during data generation.