OmniVoice Latin LoRA
LoRA adapter (r=16, step 4000) for k2-fsa/OmniVoice, trained on ~22 h of macronized Latin with restored classical pronunciation.
Usage: add "latin": "la" to LANG_NAME_TO_ID in omnivoice/utils/lang_map.py, then
python -m omnivoice.cli.infer --model k2-fsa/OmniVoice --lora_adapter <this repo, downloaded> \
--language la --text "..." --ref_audio ref.wav --ref_text "..." --output out.wav
Data: curated subset of Vox Classica (Ken-Z/Latin-Audio, CC-BY-4.0) plus Fables of Orbilius and additional recordings.