Kryton Scribe v1

Kryton Scribe v1 merges the compatible encoders of two Apache-2.0 Whisper Small checkpoints: 90% multilingual and 10% English. The multilingual decoder, tokenizer, and generation configuration are retained because the source vocabularies differ by one token.

The Transformers directory passed a CPU load and transcription test against the included Kryton Voice smoke sample, returning Hello world. Broader multilingual word-error-rate, translation, timestamp, and hallucination evaluation is still recommended.

Exact source revisions, verified hashes, blend weights, tensor counts, support-file hashes, and the output hash are in kryton-merge-manifest.json. build_model.py reproduces the merge.

Downloads last month
11
Safetensors
Model size
0.2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support