NSynth is "A large-scale and high-quality dataset of annotated musical notes". The dataset is provided as a tfrecord file. However, if you are not using tensorflow (like me) you may find it quite inconvenient to use. This is a little converter that read the tfrecord files and writes out pcm audio + json metadata as one file per sample.

Make sure you have installed libsndfile. Then install the python requirements by

pip install -r requirements.txt

to convert a tfrecord file and write standard json + audio files into a folder called NSynth-Test just run

python convert.py nsynth-test.tfrecord NSynth-Test