Golha XLS-R 300M Persian ASR Student
Fine-tuned CTC ASR student based on facebook/wav2vec2-xls-r-300m.
This checkpoint was trained on VAD-trimmed Golha spoken-word clips labeled with Gemini 3.5 Flash after QC. It is intended as a fast streaming-capable Persian ASR student and an intermediate teacher candidate for the broader Persian ASR relabeling loop.
Training Snapshot
- Base model:
facebook/wav2vec2-xls-r-300m
- Data: Golha spoken-word clips, 16 kHz FLAC, Gemini-labeled
- Training host: RTX PRO 6000 Blackwell, 10 epochs
- Best checkpoint:
checkpoint-4000
- Trainer validation: WER 26.16%, CER 9.63%
Full external eval on Golha gold-69 and FLEURS-fa was launched after upload; see repository files for double_eval_results.json when available.
