latin_whisper-large-v3-turbo

openai/whisper-large-v3-turbo fine-tuned on Vox Classica.

Results

Test set: 2,060 clips from speakers who do not appear in training (5.6 h). Decoding: beam size 5, language="latin". Before scoring, both reference and prediction are normalized: diacritics (macrons) stripped, lowercased, v→u, j→i, digits and punctuation removed.

Model	WER %	CER %
latin_whisper-large-v3-turbo (this model)	13.01	2.74
openai/whisper-large-v3-turbo (no fine-tuning)	35.55	8.03

Splits (splits.json) are included in this repository.

Usage

from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="Ken-Z/latin_whisper-large-v3-turbo")
print(asr("clip.wav", generate_kwargs={"language": "latin", "task": "transcribe"})["text"])

For audio longer than 30 seconds, add chunk_length_s=30 to the asr(...) call.

Output is lowercase Latin without macrons or punctuation, written with u and i (no v or j).

Training

• Base model: openai/whisper-large-v3-turbo
• Data: Vox Classica train split (58.2 h), speaker-held-out validation and test
• Hyperparameters: lr 1e-05, 500 warmup steps, batch 16, 3 epochs, optimizer adamw_bnb_8bit, SpecAugment
• Best checkpoint chosen by validation WER

Limitations

• Trained in restored Classical pronunciation. Expect higher error on Ecclesiastical pronunciation.
• No macrons (vowel length), punctuation or capitalization in the output.

Data credits

Vox Classica (CC-BY-4.0)