NOTICE: components, license chain and attribution
These Korean verbatim models are derived works. The effective license is the most restrictive of the components below: non-commercial / research use only, with additional AI Hub KsponSpeech terms on the Korean capability.
License chain recorded to the best of our knowledge as of 2026-08. Not legal advice. Re-verify each source's current terms before distributing or deploying, especially for any commercial use.
Teacher / init / tokenizer
- CrisperWhisper2.0_small (nyralabs/CrisperWhisper2.0_small), by Nyra Health. Used as: initialization weights for the small model, the soft-KD teacher for base and tiny, and the source of the 51,896-token tokenizer. License: Nyra Health Non-Commercial Research License. These weights inherit it. https://huggingface.co/nyralabs/CrisperWhisper2.0_small
Training data
- KsponSpeech (Korean Spontaneous Speech), AI Hub / NIA, Korea. Korean verbatim ground-truth labels (about 236,600 clips; public HF mirror DragonLine/ksponspeech). License: AI Hub Terms of Use (restricted). https://www.aihub.or.kr/
- Zeroth-Korean (kresnik/zeroth_korean). Korean read speech. License: CC BY 4.0. https://www.openslr.org/40
- Whisper (openai/whisper-base, openai/whisper-tiny). Base / tiny skeletons. License: MIT.
Summary
| Component | License | Effect |
|---|---|---|
| CrisperWhisper2.0_small (init + teacher + tokenizer) | Nyra Health Non-Commercial Research | non-commercial only |
| KsponSpeech (AI Hub) | AI Hub Terms (restricted) | restricted, verify before redistributing |
| Zeroth-Korean | CC BY 4.0 | attribution required |
| Whisper (OpenAI) | MIT | permissive |