DXnet
A CNN to perform Yamaha DX7 patch regression from audio samples.
DXnet generates Yamaha DX7 patches from real audio samples. The model provided here 'desynthesizes' sounds captured in a 2 second window @ 8Khz. The CNN regresses a vector of lenght 145, which codifies a Yamaha DX7 patch.
Requirements
To train the networks and perform patch operations, the scripts use the dx7pytorch module, a Pytorch Dataset of DX7 sounds synthesized on-the-fly, which also provides a collection of almost 30k Patches.
Training script
The training script train.py trains the model from a DX7 patch collection, synthesizing instances of 2 seconds of audio sampled at 8Khz. An ad-hoc loss function is used, which applies different weights to the patch parameters.
To reduce the computational load during training, the model provided here has been trained with a reduced patch collection, only including those which are Algorithm 2 and 5. Furthermore, only 4 MIDI notes are employed for training, using maximum velocity for each instance.
Inference script
The provided inference script, wave_to_patch_file.py opens a specified .WAV file, crops the first 2 seconds, mixes and resamples it to a single channel and then performs inference on the DXnet model, generating a patch file which can be loaded into a Yamaha DX7 using a SYSEX dump. The file is also compatible with several DX7 emulators, like DEXED.