TensorFlow implementation of the ICLR 2019 paper
GANSynth: Adversarial Neural Audio SynthesisOriginal paper
Based on following papers
Dataset
Requirements
- TensorFlow 1.13.1 with GPU support.
Usage
-
Following the paper, create a new train/valid/test 80/10/10 split from shuffled data, as the original split was divided along instrument type, which isn’t desirable for this task.
-
The original tfrecord is very large, so it takes so long time to shuffle after each epoch.
For better peformance, make tfrecord which contains the path to waveform.
wget http://download.magenta.tensorflow.org/datasets/nsynth/nsynth-train.jsonwav.tar.gz
wget http://download.magenta.tensorflow.org/datasets/nsynth/nsynth-valid.jsonwav.tar.gz
wget http://download.magenta.tensorflow.org/datasets/nsynth/nsynth-test.jsonwav.tar.gz
tar -xvf nsynth-train.jsonwav.tar.gz
tar -xvf nsynth-valid.jsonwav.tar.gz
tar -xvf nsynth-test.jsonwav.tar.gz
python make_tfrecord.py
python main.py --filenames nsynth_train_examples.tfrecord --train
python main.py --filenames nsynth_test_examples.tfrecord --evaluate