Towards data-driven physical modeling synthesis

A current research trend is the combination of physical models with machine learning. These hybrid models leverage the strength of both fields and require smaller datasets in training. Trained models are interpretable and well suited towards solving inverse problems. We present recent advancements on differentiable digital waveguide and reed modeling. The proposed models are implemented in the PyTorch framework and can extract physical parameters from audio data. The proposed models allows us to examine player interactions by retrieving player parameters from audio data.