dataloaders

Pytorch and TFRecords data loaders for several audio datasets

Datasets

1. ESC - dataset of environmental sounds

• [x] ESC Downloader
• [x] Pytorch DataSet
• [x] TFRecords Loader

2. LibriSpeech - corpus of read English speech

• [x] LibriSpeech downloader for PyTorch
• [x] PyTorch DataSet
• [x] PyTorch DataSet for TFRecord
• [x] PyTorch DataLoaders for TFRecord
• [x] TFRecords Loader
• [x] TFRecords Generator

3. NSynth - dataset of annotated musical notes

• [x] NSynth downloader and generator of *.h5py and *.tfrecord formats
• [x] TFRecord reader
• [x] PyTorch Dataset
• [x] PyTorch Dataset for TFrecord
• [x] PyTorch DataLoaders for TFRecord

4. VoxCeleb2 - human speech, extracted from YouTube interview videos

• [ ] Pytorch loader
• [ ] TFRecords loader

5. GTZAN - audio tracks from a variety of sources annotated with genre class

• [x] GTZAN Downloader
• [x] PyTorch DataSet

6. CallCenter - audio tracks with human and non-human speech

• [x] PyTorch DataSet

For validation we frequently use the following scheme:

1. Read 10 random crops from a file;
2. Predict a class for each crop;
3. Averaging results.

For this scheme we've done additional DataLoaders for PyTorch:

• DataLoader for ESC, GTZAN, LibriSpeech
• DataLoader for LibriSpeech from TfRecords
• DataLoaders for NSynth