Diffsynth - a Differentiable Musical Synthesizer in PyTorch

Synthesizer Sound Matching with Differentiable DSP @ ISMIR2021 https://hyakuchiki.github.io/DiffSynthISMIR/

Features

• Additive-subtractive synthesizer
• FM synthesizer
• ADSR envelopes, LFOs
• Chorus/flanger, reverb effects
• Parameter estimator network

To-do

• Training with perceptual loss doesn't work

Training

• p-loss model
  • python train.py experiment=only_param_h2of trainer.gpus=1
• pretrain
  • python train.py experiment=pretrain_h2of trainer.gpus=1
• resume real model
  • python train.py experiment=resume_real_h2of trainer.gpus=1 trainer.resume_from_checkpoint=[pretrain ckpt absolute path]
• resume synth model
  • python train.py experiment=resume_synth_h2of trainer.gpus=1 trainer.resume_from_checkpoint=[pretrain ckpt absolute path]

Notes

Some code was ported to pytorch from DDSP (Copyright 2019 Google LLC.).

Several features have been added since ISMIR2021 (code is more readable, chorus effect, etc.) To reproduce results from ISMIR2021, revert to Ver. May2021