| NepTrans v4, shared by DontHash, 2026. | |
| This release adapts openai/whisper-large-v3 at revision | |
| 06f233fe06e710322aca913c1bc4249a0d71fce1. The pinned Hugging Face | |
| model card declares Apache-2.0. The original Whisper project is | |
| Copyright (c) 2022 OpenAI and distributes its code and weights under MIT; | |
| the original MIT notice is retained in WHISPER_LICENSE. Changes consist | |
| of NepTrans LoRA fine-tuning and, for the CT2 package, adapter merging | |
| and CTranslate2 int8_float16 conversion. Upstream notices are retained. | |
| Model artifacts: Apache License 2.0 (LICENSE). | |
| Included helper/protocol code: MIT (CODE_LICENSE). | |
| Training data attribution: | |
| Large Nepali ASR training data set, OpenSLR-54. | |
| Copyright 2016, 2017, 2018 Google, Inc. | |
| Source: https://www.openslr.org/54/ | |
| License: Creative Commons Attribution-ShareAlike 4.0 International. | |
| The model used a curated approximately 15-hour subset of this corpus. | |
| The dataset license is preserved in OPENSLR54_LICENSE. | |
| See the model card for the requested Kjartansson et al. (2018) citation. | |
| No dataset audio, transcript, or private manifest is redistributed here. | |
| Runtime libraries are separate dependencies with their own licenses: | |
| faster-whisper (MIT), CTranslate2 (MIT), Transformers (Apache-2.0), | |
| and PEFT (Apache-2.0). They are not vendored in these model repositories. | |