Fix optional test with fairseq

[vincentqb]

• As seen here, we need to bump the python version used in example test in travis to at least 3.6.
• Fix test
• Disable test

internal doc

[tomassosorio]

Looking at for example:
pytorch text, we can see that this show with pytorch and torchtext how to build a text classification, however I do not see the point of this example in the scope of torchaudio since basically just uses the fairseq, which already have examples in its repository.

I think would be beneficial to do an example only using the functionalities from torchaudio and pytorch to do an audio classification with SpeechCommands for example. This way really showing what torchaudio is meant for and its utility.

[vincentqb]

Thanks for the feedback :) This example was a partnership, and was helpful during the hacktackon last summer.

I agree with you that it'd be nice to have reference pipelines for researchers that also showcase torchaudio's ability more specifically. What kind of example would you suggest with SpeechCommands?

[tomassosorio]

Currently, tensorflow has a tutorial, maybe would be better to use yesno dataset, I do not know the politics of using a similar example with the same dataset produced by them, and create a simple classification model, as a beginners tutorial, and then for an intermediary tutorial a simple ASR example, using pytorch implemented CTC and SpeechCommands.

[vincentqb]

@tomassosorio -- We would indeed welcome new tutorials that uses pytorch, torchaudio, and freely available datasets :) We could add it here in https://github.com/pytorch/tutorials. Is that of interest to you?

[tomassosorio]

Absolutely, I will start working on that!

[vincentqb]

Awesome! when you open a pull request https://github.com/pytorch/tutorials, make sure to ping me on it :)

[vincentqb]

btw, feel free to open an issue over there to discuss how we should turn this into a reference implementation

[vincentqb]

This test is no longer ran as part of CircleCI.