Add spoken digit dataset and baseline model

[faroit]

🚀 Feature

Given that there is a lack of small and comprehensive audio tasks, I would propose to add a speech MNIST dataset to torch audio.

Motivation

In the audio domain, we often lack small toy scenarios that would be a good equivalent to the ubiqous MNIST task.
A spoken digit dataset and model could help to sketch and try audio ML ideas.

Pitch

add either of the two:

• FSDD
• AudioMNIST

to torchaudio.datasets

Additional context

furthermore, it might be a good idea to also add a baseline model, either based on MELSpectrogram -> conv2d or using the existing wav2letter.

[vincentqb]

The two datasets sound great :) and the example you mention would make for a nice tutorial, e.g. pytorch/tutorials#1204, thoughts?

[faroit]

@vincentqb I expect a lot of overlap with a possible baseline for the speechcommand dataset. Is there a simple and lightweight ASR model that we can use for a tutorial?

[vincentqb]

oops, I updated the link in my previous message to the new audio classification tutorial we have for audio. Is that what you meant?

[jonnor]

There is now a PyTorch loader for FSDD in https://github.com/eonu/torch-fsdd

[jonnor]

There was also a pull request to add AudioMNIST, but it was closed, #84