| license: other | |
| license_name: celestial-princess-noncommercial-technical-preview-terms-v1 | |
| license_link: LICENSE | |
| library_name: GPT-SoVITS | |
| tags: | |
| - text-to-speech | |
| - voice-cloning | |
| - synthetic-voice | |
| # CelestAI | |
|  | |
| <audio controls src="samples/smoke-test.wav"></audio> | |
| Unofficial fan-made GPT-SoVITS voice model of Princess Celestia. | |
| Trained to sound like Nicole Oliver’s performance. Not affiliated with, endorsed by, or connected to Nicole Oliver, Hasbro, or anyone who actually owns the character. | |
| This is a non-commercial technical preview. Every single output must be clearly labeled as AI-generated synthetic speech and unofficial fan work. Do not pretend the audio is real, an endorsement, or anything official. | |
| If you host this, keep the repo private while uploading and turn on manual access requests before making the page public. Model card tags don’t enforce access control. | |
| ### What’s inside | |
| - Two fine-tuned inference checkpoints | |
| - One model-generated reference WAV + original text | |
| - The stable inference profile that worked best | |
| - A smoke-test sample | |
| - SHA256 checksums | |
| No episode audio, no cleaned source clips, no transcripts, no training data, no recovery checkpoints. Just the model. | |
| Trained on ~29 minutes of manually reviewed dialogue. Source audio is not included. This package makes no claim that consent or a distribution license was obtained from the performer or rights holders. | |
| ### Runtime | |
| Use GPT-SoVITS `20250606v2pro` at commit `d7c2210da8c013e81a94bfc7b811a477c99fd506` with pretrained revision `336b2ec4e8d4ac74740798dd40af44e74659ecaf`. | |
| Verify `checksums.sha256` before loading anything. PyTorch checkpoints can run code on load — only use files from the authenticated release that match the published hashes. | |
| Default profile and seed are in `profile.json` and `manifest.json`. | |
| Read `ACCEPTABLE_USE.md` before you do anything with it. | |
| To use CelestAI, open GPT-SoVITS 20250606v2pro, select `weights/model.ckpt` as the GPT model and `weights/model.pth` as the SoVITS model, then use `reference/reference.wav` as the reference audio and paste the exact prompt from `reference/reference.txt`. Apply the inference settings recorded in `profile.json`, enter your text, and use seed `314159` for the selected voice profile. Verify `checksums.sha256` before loading the model, and clearly label shared audio as AI-generated fan work. | |