Thank you, @BenoitWang , for your PR contribution! t's great to see this valuable feature being added to the FastSpeech model.
I'm running the recipe, but I still have some comments to share with you:

1. Recipe Tests:
   The recipe tests are currently failing. Please run the following command to check:

python -c 'from tests.utils.recipe_tests import run_recipe_tests; print("TEST FAILED!") if not(run_recipe_tests(filters_fields=["Dataset"], filters=[["LJSpeech"]], do_checks=False, run_opts="--device=cuda")) else print("TEST PASSED")'

I encountered this error:

RuntimeError: These keys are still unaccounted for in the data pipeline: phonemes

2. FastSpeech2.py:
   In the speechbrain/lobes/models/FastSpeech2.py file, please consider the following improvements:

• Add a docstring example for the LossWithAlignment class.
• Describe the arguments and add a docstring example for the ForwardSumLoss class.
• Describe the arguments and add a docstring example for the BinaryAlignmentLoss class.
• Add a docstring example for the maximum_path_numpy function.
• Consider whether it makes sense to write a docstring example for the TextMelCollateWithAlignment class.

3. interfaces.py:
   In the speechbrain/pretrained/interfaces.py file, I noticed the addition of FastSpeech2InternalAlignment. Is there are reason for that? Ideally, we should strive to reuse the existing FastSpeech2 interface if possible and convenient.

4. text_to_sequence.py:
   For the g2p_keep_punctuations function in speechbrain/utils/text_to_sequence.py, please describe the arguments and add a docstring for clarity.

Hi @mravanelli , thanks for the comments! I've fixed the recipe test and added the necessary docstrings and examples.
However, for LossWithAlignment, since it's a summary of the other losses, an example takes too many inputs and seems repeated because an example can be found in each component loss.
As for the new interface, I created a new one mainly because the outputs of the models are different, the internal alignment model outputs the alignments in addition. The inference processes are very different as well, for example the original one needs the spn prediction while the new one focuses on keeping the punctuations during g2p.

Thank you @BenoitWang. I did the last tests and everything seems to work properly. Thank you for this new model that improves SpeechBrain's TTS systems currently available.