Added Google Speech Commands V2 Dataset

[harritaylor]

Summary

As title, added google speech commands dataset, with some choice designs for preprocessing as described in the next section.

Proposed Changes

• Added speech commands dataset
• Dataset partitioning into train / test using testing_list.txt from the dataset download
  • validation_list.txt is still provided for use to use sampler in the dataloader if train/valid/test is desired
• Using include keyword for defining custom subset of dataset. All other words in the dataset are marked as unknown (not sure if this adheres to the original split, but couldn't find more information about this)
• Using silence keyword to include samples from _background_noise_ folder

Discussion

1. silence: currently the DataSet simply points to samples in the _background_noise_ folder. These samples are 1min long, whereas the speech commands are 1sec long. My current workaround is to use RandomCrop with 16,000 samples, which deals with this issue nicely. I don't think it would be efficient to chop up and store up 1 second clips of the silence clips.

[hagenw]

Cool, thanks for your hard work. I added just one small comment.

And if you like, feel free to add your name in LICENSE under the Contributors: section.

[harritaylor]

Cool, thanks!