Add Galician model v0.1.3

[zuazo]

A Galician model trained based on the process described by Francis M. Tyers here: https://arxiv.org/abs/2105.04674

• Deep Speech model trained on Common Voice 12.
• Hyperparameter sweep to find the best model configuration: LR=0.00001, dropout=0.2 and SpecAugment.
• LM trained on OPUS, Wikipedia, and SLI GalWeb 1.0.

The accuracy:

Test Corpus	WER	CER
Common Voice	16.4%	6.8%

The models can be downloaded from here: https://aholab.ehu.eus/~xzuazo/models/Galician%20STT%20v0.1.3/