Commercial use and fine-tuning permissions for Ria and NSF-HiFiGAN with DiffSinger

[tobyapi]

Hi OpenVPI team,

I’m using OpenVPI DiffSinger in a local singing synthesis application. I would like to clarify commercial-use permissions for the complete setup, including Ria and the vocoder.

My current setup uses:

• Ria v0.4-lynxnet, including its acoustic, duration, pitch, and variance models.
• The fine-tuned NSF-HiFiGAN vocoder distributed by Ria’s author, based on the OpenVPI 2024.02 model.

I want to use generated vocals in songs sold as digital downloads, distributed through royalty-paying streaming services, and included in monetized videos. I would release finished audio only, without redistributing model weights or offering a hosted synthesis service.

I have read DiffSinger’s Apache 2.0 license, the community vocoder’s CC BY-NC-SA 4.0 terms, and Ria’s usage rules. I’m unsure which permissions cover commercial music production and which conditions apply to the resulting songs.

Could you clarify the following, or direct me to the relevant rights holders?

1. Commercial songs using Ria: Do Ria’s current terms permit the commercial uses described above? Is crediting “Ria” or “狸安” sufficient, or are additional permission, fees, or conditions required?
2. Fine-tuning Ria: Assuming access to compatible training checkpoints, is it permitted to fine-tune Ria’s acoustic, duration, pitch, or variance models using my own authorized recordings to adapt the voice’s timbre or singing style? May I then use the resulting models to produce commercial songs? Do Ria’s restrictions on voice-changing software also affect this type of model adaptation?
3. Vocoder permissions: Does the upstream NSF-HiFiGAN license permit using its pretrained weights to produce commercial songs when the weights remain local? For the version fine-tuned by Ria’s author, would I need additional permission from that author? Would further fine-tuning on my own recordings change the applicable conditions?
4. Conditions on the audio: If commercial production is permitted, what attribution or ShareAlike requirements, if any, apply to the generated vocal tracks and finished songs, as distinct from the model weights?
5. Training an independent voice: If I train all singer models from scratch on my own authorized recordings, use no Ria-derived weights, and use the official MIT-licensed BigVGAN v2 vocoder, are there any OpenVPI-specific conditions beyond Apache 2.0 that apply to commercial releases?
   Could you also confirm whether OpenVPI handles permissions for Ria, or whether I should contact RibosomeK separately? If any of these uses require a separate commercial license or written permission, please let me know the appropriate contact.

I found the related openvpi/vocoders issue #4, which concerns PC-NSF-HiFiGAN. My current setup uses the earlier NSF-HiFiGAN version described above.

Thank you for your help.