Model description
Training and evaluation data
This model was fine-tuned on a mixture of publicly available Serbian speech corpora, including:
- Mozilla Common Voice 24.0, evaluated on CV test (sr)
- FLEURS Serbian
- ParlaSpeech-RS (subset of the full dataset)
- Additional Serbian corpora used in the training pipeline
Training procedure
- Epochs: 9
- Batch size: 32 / 20
- Optimizer: AdamW
- LR: 6e-5 with warmup (50 steps) + cosine decay to min_lr = 1e-7
- Mixed precision: bfloat16 (fp32 in the final epoch)
- SpecAugment: frequency + time masking
- Sampling: weighted sampling across datasets
Training results
Table with columns: Epoch, Train loss, CV WER| Epoch | Train loss | CV WER |
|---|
| 1 | 0.333 | 0.1614 |
| 2 | 0.344 | 0.1278 |
| 3 | 0.251 | 0.1112 |
| 4 | 0.202 | 0.1032 |
| 5 | 0.167 | 0.0934 |
| 6 | 0.138 | 0.0790 |
Evaluation Metrics
- WER (normalized) on Common Voice 24.0 Serbian test: 7.09%
- Text normalization used for WER:
- punctuation removed
- lowercased
- Cyrillic → Latin conversion
- numbers converted to words