Instructions to use harrrshall/kriti with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use harrrshall/kriti with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("harrrshall/kriti") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
| kriti includes weights derived from the ai4bharat nepali indicconformer. | |
| base model: | |
| https://huggingface.co/ai4bharat/indicconformer_stt_ne_hybrid_ctc_rnnt_large | |
| revision cd09ba7720f3b17d259f6bfd03e1463bc5ba517d | |
| runtime: | |
| https://github.com/AI4Bharat/NeMo | |
| revision 8dce88cf8e94963e2033c3137f7b9993b51db88a | |
| the upstream model card declares the checkpoint under the mit license. kriti | |
| retains its conformer encoder, prediction network, and nepali rnnt weights. the | |
| kriti release adds language-specific pruning, runtime embedding compaction, a | |
| 1,025-parameter acoustic danda head, artifact verification, and public | |
| reproduction code. | |
| openslr 54, openslr 43, fleurs, and indicvoices source material retain their | |
| respective licenses and access terms. their speech, transcripts, speaker | |
| metadata, and row manifests are excluded from both the source repository and | |
| the model repository. | |
| benchmark checkpoints retain the authorship, ownership, model cards, and terms | |
| published by their respective maintainers. references.md records their exact | |
| role in the released research record. | |