whisper-medium_ro-80mel β Romanian Fine-tune
Fine-tuned openai/whisper-medium
on a merged Romanian corpus (~445h). Input: 80-mel spectrograms.
Performance (beam=5, full val)
| Metric |
Value |
| WER |
0.0656 |
| CER |
0.0228 |
| Evaluated on 25,348 validation samples (52.75h). |
|
Training
| Parameter |
Value |
| Base |
openai/whisper-medium |
| GPU |
NVIDIA A100 80GB PCIe |
| Features |
80-mel |
| Effective batch |
128 (16Γ8) |
| Optimizer |
adamw_torch |
| LR / scheduler |
4e-06 / cosine |
| Warmup |
500 |
| Max epochs |
8 |
| Precision |
BF16 |
| Selection |
eval_loss (early stop 5) |
| Date |
2026-09-21 |
Dataset
| Dataset |
Samples |
Hours |
| Echo |
163,050 |
374.33h |
| VoxPopuli |
27,060 |
85.20h |
| RODigits |
15,389 |
37.55h |
| RO-Smart (TTS) |
10,572 |
17.28h |
| CommonVoice-25 |
19,850 |
22.15h |
| FLEURS |
4,161 |
13.73h |
| USPDATRO |
2,637 |
4.31h |
| TOTAL |
242,719 |
554.55h |
Notes
- decoder_start_token_id=50258 (SOT); positional label masking; begin_suppress_tokens=[220].
- Checkpoint selection on eval_loss; final WER via standalone beam=5 generate.