Multi-language capabilities for MatchaTTS

[fzalkow]

This pull request adds multi-language support to MatchaTTS by concatenating language embeddings to the encoder and decoder inputs.

Speaker and Language Disentanglement:
When training with only a few monolingual speakers, speaker and language IDs are highly correlated. Therefore, it is likely (though not tested) that the model may have difficulty disentangling them. With multilingual speakers or a sufficiently large number of speakers, the model can better separate speaker and language information.

How to Use:
Specify the number of languages using the n_langs key in the data config and annotate your data CSV as follows:

filepath	spk	lang	text

Backward Compatibility:
If you set n_langs: 1, the system behaves as before this pull request, with one exception: both speaker and language embeddings are now L2-normalized (if used). In my experience, this helps maintain a balance between speaker and language information during training.

Note:
This pull request does not include any text processing functions for additional languages. You will need to ensure that your text preprocessing pipeline supports the languages you intend to use.

[scott-parkhill]

Would this PR allow us to train a multi-lingual model on typologically similar languages to potentially reduce the amount of training data we would need for these languages? I.e. if we have a lot of training data for one language in the family, and there is a topologically similar language for which we do not have much data, could the contents of this PR allow us to train a voice leveraging the model trained with a larger dataset?

[fzalkow]

Would this PR allow us to train a multi-lingual model on typologically similar languages to potentially reduce the amount of training data we would need for these languages? I.e. if we have a lot of training data for one language in the family, and there is a topologically similar language for which we do not have much data, could the contents of this PR allow us to train a voice leveraging the model trained with a larger dataset?

I have not tried this for this particular PR, but for other models with similar techniques, and there was a cross-language benefit. So, my guess is that the answer to your question is yes.