exp_009_small_cv24_vanilla
Fine-tuned openai/whisper-small on Common Voice Pashto 24 (train+dev splits, ~113h combined).
Evaluated on the held-out test split (13,791 samples, ~20.72h).
Results
| Metric | Value |
|---|---|
| WER | 24.89% |
| CER | 8.075% |
| Step | 8500 |
| Epoch | 5.85 |
Best checkpoint: checkpoint-8500
Training details
- Base model: openai/whisper-small
- Dataset: ihanif/common-voice-pashto-24 — train+dev splits (~113h)
- Eval/test set: test split (held out, never seen during training)
- Framework: HuggingFace Transformers + Seq2SeqTrainer