FM Sound Match Experiment

This experiment compares several different algorithms for estimating parameters for an FM synthesizer. The goal is to be able to select synthesizer parameters in order to replicate a target sound as closely as possible. This is called sound matching. We’ll run this experiment using the open-source Dexed VST emulation of the Yamaha DX7. Dexed can be dowloaded for free here.

Through this example we will use SpiegeLib to:

• Program and generate sounds from a VST synthesizer

• Generate datasets for deep learning and evaluation

• Train deep learning models

• Perform sound matching using deep learning and genetic algorithms

• Evaluate results

If you want to follow along or recreate any part of this experiment, make sure you have SpiegeLib and RenderMan installed. See installation instructions. And download Dexed.

If you want to jump ahead and hear the results, check out the audio results page

All code is available as Python notebooks on the project github page. The trained models from this experiment are also available in the git repo. All datasets generated and used in this experiment are also available online: https://doi.org/10.5281/zenodo.3722784.

Experiment Sections

• Synthesizer Configuration
• Dataset Generation
• Train Deep Learning Models
• Sound Match Deep Learning Models
• Sound Match Genetic Algorithms
• Evaluation
• Audio Results