MFA Restructure & Environment Bug Fixes

[Merakist]

✨ Description

• Address fs2 preprocessing exception due to mismatched package versions
• Refactor MFA file structure into /pretrained and modify control logic in fetching and managing MFA files

🚧 Related Issues

• Issue [BUG]: Fix for AssertionError When Running FastSpeech2 Preprocessing （run.sh --stage 1） #118, error in training from LJSpeech dataset, arising from outdated file paths for lexicon dictionary.

• Issue [BUG]: TypeError in pad_center() During FastSpeech2 Preprocessing Due to librosa Version Mismatch #119, due to package manager installing librosa 0.10+, which deprecated audioread support (See librosa 0.10 documentation)

• Further testing identified additional environment issues arising from newer dependency versions causing errors. After isolation testing, diffsptk is confirmed to cause the error. More specifically, the dependency vector-quantize-pytorch. The package manager fetches vector_quantize_pytorch 1.12.16 by default (notice the underscores instead of hyphens in the package name) instead of vector-quantize-pytorch. Therefore, the version should also be specified for this package to be no more than 1.12.5 for now until a change in package manager behavior.

👨‍💻 Changes Proposed

• Modifying directories in preprocessors/ljspeech.py to change the saving directory for MFA files from the root folder to /pretrained to comply with file management convention for Amphion system; additionally, link lexicon directory to provided lexicon file to avoid redundant files.
• Rewrite prepare_mfa.sh and modify run.sh for a more robust logic in fetching and managing MFA files; removing the redundant section for downloading LJSpeech lexicon.
• Modifying env.sh to specify librosa version to 0.9.1 and vector-quantize-pytorch to 1.12.5

🧑‍🤝‍🧑 Who Can Review?

@lmxue @RMSnow

🛠 TODO

Potential Consideration: A few issues seem to arise from preparing environments, one consideration could be to freeze versions for certain packages for better stability @RMSnow.

✅ Checklist

• Code has been reviewed
• Code complies with the project's code standards and best practices
• Code has passed all tests
• Code does not affect the normal use of existing features
• Code has been commented properly
• Documentation has been updated (if applicable)
• Demo/checkpoint has been attached (if applicable)

[RMSnow]

Nice work!

[lmxue]

Have you tested the MFA processing pipeline with the updated code?

[Merakist]

Have you tested the MFA processing pipeline with the updated code?

Yes, the modified MFA processing pipeline ran without errors occurring.

[lmxue]

LGTM. Thanks for your contribution.