Multivocoder Update: A substantial infrastructure upgrade proposed as a version improvement.

[blaisewf]

We propose this update to modernize the code infrastructure. Currently, it will be a public experiment, gathering strong feedback from the community. If the results are deemed optimal, the version will be released only after completing the training phase of the base pre-trained models and the final optimization.

Objectives

• [x] Implement BigVGAN (V2)
• [x] Implement BigVSAN
• [ ] Implement Vocos
• [ ] Optional: Implement HiFiSAN
• [ ] Idea: Optimize BigVGAN and BigVSAN activation to reduce high resource consumption by Snake
• [ ] Idea: Introduce automatic f0 prediction (SVC)
• [ ] Train pre-trained models on all the vocoders (VCTK)

Updates

• 08/06/2024: Start of development...
• 30/07/2024: Experimental testing phase
• 28/10/2024: Back to the interest in the project...?

You should not view these proposed changes as a reliable solution. Instead, consider them as an unstable and inadequately developed experiment.