Differentiable Synthesis Libraries

Differentiable Synthesis Libraries#

There exist several open-source code implementations of differentiable DSP components for creating and working with DDSP synthesizers.

This includes implementations of the differentiable sinusoidal modelling synthesizer used in the work by Engel et al. and synthesis libraries modelled on modular synthesizers.

DDSP TensorFlow#

Full repository

The official GitHub repository for Engel at al.’s DDSP.

DDSP PyTorch#

Full repository

Implementation of the DDSP model using PyTorch by ACIDS IRCAM.

torchsynth#

Full repository

A GPU-enabled modular synthesizer library written in PyTorch. Provides of a number of different synthesis modules that are inspired by classic modular synthesizers that can be combined to create larger synthesizer voices. Achieves over 16k times faster than realtime GPU performance for fast synthetic dataset generation. [TST+21]

SynthAX#

Full repository

A modular synthesizer based on torchsynth built with JAX. Achieves over 80k times faster than realtime generation speeds. [CS23]

dasp#

Full repository

Differentiable audio signal processors in PyTorch. Includes differentiable signal processors such as distortion, EQs, compressors, reverb, and stereo effects. Not focused on audio synthesis, but relevant to many of the concepts presented in this tutorial.

References#

CS23

Manuel Cherep and Nikhil Singh. Synthax: a fast modular synthesizer in jax. In Audio Engineering Society Convention 155. May 2023. URL: http://www.aes.org/e-lib/browse.cfm?elib=22261.

TST+21

Joseph Turian, Jordie Shier, George Tzanetakis, Kirk McNally, and Max Henry. One Billion Audio Sounds from GPU-enabled Modular Synthesis. In Proceedings of the 23rd International Conference on Digital Audio Effects. 2021.