GitLab

* Add wav2vec2 ASR German pretrained model from voxpopuli

In Wav2Vec2 ASR pipelines, the `get_model` method performs on-the-fly model
surgery to remove unused dimensions common to all the Wav2Vec2 model trained
with fairseq.

In VoxPopuli, there seems to be an extra dimensions introduced due to
some issue in the preprocessing stage.

For example, the label `1` shows up in the training dataset of German (1 out of 16M),
English (1 / 28M), Spanish (1 / 9.4M), Romanian (1 / 4.7M) and Polish (6 / 5.8M).

This code changes will allow the customization of excluded dimensions for such cases.

- Make the test support other languages
- Fetch tetst asset on-the-fly

It seems that this is no longer necessary for recent macOS .

* [BC-breaking] Remove unused dimension from pretrained Wav2Vec2 ASR

The Wav2Vec2 ASR pretrained weights originated from fairseq have
extra dimension that have nothing to do with the ASR task.

https://github.com/pytorch/fairseq/blob/c5ff181125c7e6126b49a85e5ebdd5f5b6a07914/fairseq/data/dictionary.py#L18-L37

which is masked during the loss computation as

https://github.com/pytorch/fairseq/blob/c5ff181125c7e6126b49a85e5ebdd5f5b6a07914/fairseq/criterions/ctc.py#L126-L128

This change removes it.

* Use '-' for blank token representation.

This is follow up after https://github.com/pytorch/builder/pull/823 that gets rids of `feature` and migrate it to `run_constrained`

If `torch.hub` is never used, the cache directory "~/.cach/torch/hub/checkpoints/" does not exist,
and the attempt to download DeepPhonemizer checkpoint there would fail.

This commit fixes it by calling `os.makedirs(directory, exit_ok=True)`.

1. Override the return type so that Sphinx shows the exported symbols.
   (output model types and input torch.nn.Module)
2. Tweak docs for Tacotron2TTSBundle interfaces
3. Fix for HUBERT_ASR_XLARGE