TITAN: A Versatile, Robust, and High-Quality Pretrained Model for Retrieval-based Voice Conversion (RVC) Training

Overview

TITAN is a state-of-the-art pretrained model designed for Retrieval-based Voice Conversion (https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/) training. It offers a robust solution for transforming voice characteristics from one speaker to another, providing high-quality results with minimal training effort.

Model Details

Titan-Medium

• Training Environment: Utilized a RTX 3060 TI on Applio v3.1.1 (https://github.com/IAHispano/Applio), employing a batch size of 8 over a span of 3 weeks.
• Iterations (48k): 1018660 Steps and 530 Epochs
• Iterations (40k): 1010588 Steps and 467 Epochs
• Iterations (32k): 1001469 Steps and 463 Epochs
• Sampling rate: 48k, 40k, 32k
• Fine-tuning Process: RVC v2 pretrained with pitch guidance, leveraging an 11.15-hour dataset sourced from Expresso (https://arxiv.org/abs/2308.05725) also available on datasets/blaise-tk/TITAN-Medium.

Samples

Tests performed with a premature ckpt at ~700k steps doing all tests under the same conditions.

Titan-Large

• Details forthcoming...

Collaborators

We appreciate the contributions of our collaborators who have helped in the development and refinement of TITAN.

• Mustar
• SimplCup
• UnitedShoes

Beta Testers

We extend our gratitude to the beta testers who provided valuable feedback during the testing phase of TITAN.

• SimplCup
• Leo_Frixi
• Light
• SCRFilms
• Ryanz
• Litsa_the_dancer

Citation

Should you find TITAN beneficial for your research endeavors or projects, we kindly request citing our repository:

@article{titan,
  title={TITAN: A Versatile, Robust, and High-Quality Pretrained Model for Retrieval-based Voice Conversion (RVC) Training},
  author={Blaise},
  journal={Hugging Face},
  year={2024},
  publisher={Blaise},
  url={https://huggingface.co/blaise-tk/TITAN/}
}