AdpatToAIera/oreero-whisper-sk

Oreero — Slovak Whisper models for WhisperKit (Core ML)

Core ML conversions of two Slovak-fine-tuned Whisper models, packaged in the WhisperKit folder layout so iOS/macOS apps can download and run them fully on-device.

Folder	Base model	Architecture	Slovak WER (FLEURS / CommonVoice 21)*
oreero_whisper-large-v3-turbo-sk	NaiveNeuron/whisper-large-v3-turbo-sk	large-v3-turbo (809M)	6.4 % / 13.2 %
oreero_whisper-large-v3-sk	NaiveNeuron/whisper-large-v3-sk	large-v3 (1.54B)	5.5 % / 11.6 %

* Numbers reported by the model authors (SloPal, arXiv:2509.19270). Both models are Slovak specialists — they trade other languages' quality for the best published Slovak accuracy.

Conversion

Converted with whisperkittools (whisperkit-generate-model), unmodified weights, default precision. Each folder contains MelSpectrogram.mlmodelc, AudioEncoder.mlmodelc, TextDecoder.mlmodelc.

Credits & license

All credit for the fine-tunes goes to NaiveNeuron and the SloPalSpeech project (2,806 h of Slovak parliamentary speech). Base Whisper by OpenAI. This repo only re-packages the weights for Core ML. License: MIT, same as the source models.

Published for Oreero — a local-first meeting dictaphone where audio never leaves the device.