Chore/small changes to wav2vec2 finetuning

[saattrupdan]

This PR implements the following:

1. Fixes an error related to decoding, where duplicate characters would be collapsed (e.g., "hotellet" -> "hotelet"). This was due to the removal of pad tokens during decoding, which we shouldn't do, since they act as CTC boundaries.
2. Small changes (config changes, some logging)

With this setup, we get quite close to reproducing the original setup. More precisely, we arrive at around 16 WER on Common Voice 9.0, which is not too far away from the SOTA 11 WER, and seems to mainly be due to overfitting. Here are the training plots:

Notably, I sample NST-da and CV9.0 equally, so we go through a lot of Common Voice samples in 120k steps! The sampling ratios should probably be changed, and the masking probabilities should probably be increased.

In any case, I think we're probably close enough to a reproduction to trust the framework and start training models on more data, in any case.

Oh, and I tried freezing the parameters, and also to increase the warmup time. Freezing the parameters didn't help at all, and the increased warmup time might have helped a little bit.

[sorenmulli]

Awesome with these results! Good job making a much nicer ASR framework give good results :D
lol @ such good news are in a PR titled "chore/small changes ..."
Just request review when ready :)

[saattrupdan]

Awesome with these results! Good job making a much nicer ASR framework give good results :D lol @ such good news are in a PR titled "chore/small changes ..." Just request review when ready :)

Alright, ready for review now. I'm currently optimising the training hyperparameters and I'll just leave the config change for another PR.

[AJDERS]

LGTM!