PC-DDSP-LoFiVocoder Model Family
Overview
Welcome to the official Hugging Face repository for the PC-DDSP-LoFiVocoder Model Family, a collection of vocoder models designed for DiffSinger voicebanks for use in OpenUTAU. This project provides different model checkpoints, reflecting different stages of the training process, offering users flexibility in selecting the version that best suits their needs. This vocoder aims to not be realistic but rather give a "robotic" aesthetic to the output, also aims to be pretty fast, allowing quick CPU inference.
This repository was last updated on October 13, 2025
All versions are available for download as ZIP files, including the necessary model weights, configuration files, and associated documentation.
Difference between version
- Latest: This will always the latest revision of the vocoder, trained up to October 13, 2025 (This version changes the Vocoder´s name to LoFiVocoder, please update the name from the old release to be usable)
- Ver A: Based off the latest checkpoint trained up to August 10, 2025
- Ver B: Based off an earlier checkpoint trained up to August 10, 2025, use this one if you want a slightly more robotic-ish output
Changelog:
- October 13, 2025: Version 2 release, a full retrain of the model, trained further than the original release.
- August 10, 2025: Initial Release of PC-DDSP-LoFiVocoder.
Ethical Considerations
This model is distributed under the CreativeML Open RAIL-M License, which promotes responsible AI use. Please adhere to the following:
- Use the model only for lawful purposes and avoid harmful applications (e.g., exploitation, defamation, or generating false information—see LICENSE.md for full restrictions).
- Include attribution to the original resources and this repository in any derivative works or redistributions.
Attribution
When using or redistributing this vocoder in your voicebanks, please credit the author usamireko and include both the LICENSE.md and NOTICE.md files. A suggested citation is:
"PC-DDSP-LoFiVocoder by usamireko, trained using resources from Scarfmonster/HiFiPLN (MIT), VocalSet (CC-BY 4.0), Cantoría Dataset (CC-BY 4.0), and a private dataset by Spoopy☆Ace/SpoopyAce. Available at https://huggingface.co/usamireko/PC-DDSP-LoFiVocoder."
Known Issues
- GPU rendering seems to not work properly on some instances, CPU usage encouraged (Its small enough that there´s barely a performance penalty)
- If a voicebank with a custom vocoder (dsvocoder folder), apparently makes it ignore PC-DDSP-LoFiVocoder
Both of these are getting investigated
Training Resources
This model was developed using the following datasets and codebases:
- Code: Based on Scarfmonster/HiFiPLN, licensed under the MIT License, a community vocoder framework for DiffSinger.
- VocalSet Dataset: DOI: 10.5281/zenodo.1442513, licensed under Creative Commons Attribution 4.0 International (CC-BY 4.0), provided by Julia Wilkins et al. at Northwestern University.
- Cantoría Dataset: DOI: 10.5281/zenodo.5878677, licensed under Creative Commons Attribution 4.0 International (CC-BY 4.0), provided by Helena Cuesta et al. at Universitat Pompeu Fabra.
- Private Dataset: Supplied by Spoopy☆Ace/SpoopyAce with explicit permission.
For detailed licensing terms and acknowledgments, refer to the LICENSE.md and NOTICE.md files included in the ZIP archives.
License and Legal Notices
This model is released under the CreativeML Open RAIL-M License, which grants permissions for use, modification, and distribution while imposing use-based restrictions to ensure responsible AI practices. Key points include:
- No warranties or guarantees are provided; use at your own risk.
- Redistribution must include the license and notice files.
- See LICENSE.md for the full terms and Attachment A for restricted uses.
The NOTICE.md file contains specific attributions to the training resources and contributors.
Contributing and Support
This is a community-supported project. For feedback, issues, or contributions:
- Open an issue on this Hugging Face page.
Thank you for using PC-DDSP-LoFiVocoder!