| # 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 | |
| |
