Model trained from Meta OmniASR CTC 1B on the three languages.
Training result
- Best checkpoint: step 5000
- Speaker-held-out internal WER: 24.6375%
- Internal validation CTC loss: 46.6428
- Training: 5,000 optimizer steps on an NVIDIA A100 80GB
- Data: 29,308 training utterances / 164.28 hours
- Held-out development: 4,655 utterances / 24.64 hours, speaker-disjoint from training
The internal WER above is a development diagnostic. See artifacts/validation_history.csv and artifacts/training_summary.json for the full trajectory.
Files
- best/final/model.pt: final best inference weights
- best/model.yaml: fairseq2 model metadata
- resume/step_5000/: complete model, optimizer, trainer, and data-reader resume state
- config/ctc-1b-full.yaml: reproducible training configuration
- artifacts/: validation history, raw safe metric JSONL, summary, and evaluator guard patch