okdohyuk/whisper-small-ko-coreml

whisper-small-ko (WhisperKit Core ML)

Korean fine-tune of openai/whisper-small (revision 973afd24965f72e36ca33b3055d56a652f456b4d), converted to Core ML for WhisperKit with whisperkittools 84f77a83c8f530022ae55fbb1a64b3351ef63c7a. Used by the Nogmi app for on-device Korean transcription. Folder: nogmi_whisper-small-ko-v1/.

Training data

All data is publicly licensed for redistribution of derived models and commercial use:

• FLEURS (Conneau et al., 2022), Google — google/fleurs@70bb2e84b976b7e960aa89f1c648e09c59f894dd, CC-BY-4.0 (https://huggingface.co/datasets/google/fleurs)
• Zeroth-Korean (OpenSLR 40), Atlas Guide Inc. / Gridspace — Bingsu/zeroth-korean@bd173fe2c8ed0dccd47acb4eda77542593651622, CC-BY-4.0 (https://www.openslr.org/40/)
• Mozilla Common Voice 17.0 (ko), via HF mirror fsicoli/common_voice_17_0 — fsicoli/common_voice_17_0@8262c16bf297c87a9cd88c51997c4758ed7a8ba2, CC0-1.0 (https://commonvoice.mozilla.org)

Common Voice terms: you must not attempt to determine the identity of speakers in the dataset. No KsponSpeech / AIHub data and no user recordings were used. Training: full fine-tune, lr 2.5e-06, 150 steps, batch 16×1, 18% multilingual replay (FLEURS).

Evaluation (held-out, never trained on; CER/WER in %, NFKC, punctuation and spaces removed for CER)

set	original CER	this model CER	original WER	this model WER
fleurs_ko_test	7.596	5.848	21.191	14.927
cv17_ko_test	14.438	9.472	34.658	25.525
zeroth_ko_test	10.255	6.75	33.609	25.132
fleurs_en_test	3.125	95.329	6.913	139.061
fleurs_ja_test	13.348	112.419	88.586	445.455

Original numbers were re-measured with the same script on the same rows (greedy decoding). Core ML vs PyTorch CER delta: fleurs_ko_test +0.0pp, cv17_ko_test +0.0pp.

The long-form (meeting-like) regression check was waived for this release: no internal long-form set was available.

License

Model weights: Apache-2.0 (as the base model). Attribution for CC-BY-4.0 data is given above and in NOTICE.md.