Hyperparameters — Whisper ATC Fine-tune (Run 9)
Model
| Key |
Value |
| Base model |
openai/whisper-large-v3 |
| Architecture |
Whisper Large v3 |
| d_model |
1280 |
| Encoder layers |
32 |
| Decoder layers |
32 |
| Encoder attention heads |
20 |
| Decoder attention heads |
20 |
| Mel bins |
128 |
Training
| Key |
Value |
| Optimizer |
AdamW (bitsandbytes 8-bit) |
| Learning rate |
1e-05 |
| LR scheduler |
Linear |
| Warmup ratio |
0.05 |
| Adam β₁ / β₂ / ε |
0.9 / 0.999 / 1e-8 |
| Weight decay |
0.01 |
| Per-device train batch size |
1 |
| Per-device eval batch size |
8 |
| Gradient accumulation steps |
16 |
| Effective batch size |
16 |
| Gradient checkpointing |
Yes (use_reentrant=False) |
| Mixed precision |
fp16 |
| Max grad norm |
1.0 |
| Max epochs (configured) |
30 |
| Early stop patience |
7 epochs |
| Label smoothing |
0.0 |
| Freeze encoder |
No |
| Seed |
42 |
Data Sources
| Source |
Role |
Size |
| axite_all.json |
SG military ATC synthetic (4 voices + human) |
~15,716 |
| deepdml/conversations |
Real Singapore Changi ATC VHF radio |
~1,443 |
| mnsc-part1-test |
MNSC SG-accented read speech |
~3,000 |
Augmentation
- Gaussian noise (p=0.4, amplitude 0.001–0.015)
- Time stretch (p=0.3, rate 0.9–1.1)
- Random silence padding (p=0.5, 0–0.7s each end)
- BandPassFilter (p=0.75, 300–3400 Hz, VHF radio simulation)
- Clip (p=0.2, ±0.8)
- Mp3Compression (p=0.3, 32–64 kbps)
- SpecAugment: FrequencyMasking(freq_mask_param=27) + TimeMasking(time_mask_param=100, p=0.05)
Early stopping
| Key |
Value |
| Metric |
WER (lower is better) |
| Stopped at |
Step 21185 / Epoch 19 |
| Patience |
7 epochs |
Results
| Epoch |
Eval loss |
WER |
| 1.0 |
0.0838 |
11.46% |
| 2.0 |
0.0550 |
4.28% |
| 3.0 |
0.0406 |
2.79% |
| 4.0 |
0.0417 |
6.58% |
| 5.0 |
0.0381 |
5.46% |
| 6.0 |
0.0372 |
3.27% |
| 7.0 |
0.0375 |
1.39% |
| 8.0 |
0.0381 |
5.52% |
| 9.0 |
0.0188 |
0.83% |
| 10.0 |
0.0202 |
0.84% |
| 11.0 |
0.0185 |
1.05% |
| 12.0 |
0.0189 |
0.82% ← best |
| 13.0 |
0.0189 |
0.95% |
| 14.0 |
0.0202 |
1.19% |
| 15.0 |
0.0206 |
0.91% |
| 16.0 |
0.0191 |
1.16% |
| 17.0 |
0.0169 |
1.12% |
| 18.0 |
0.0176 |
1.19% |
| 19.0 |
0.0185 |
1.19% |
Best checkpoint: training/output_run9/checkpoint-13380 (epoch 12, WER 0.82%)
Output
| Key |
Value |
| Best HF checkpoint |
training/output_run9/best/ |
| CTranslate2 model |
training/saved_models/ct2_run9/ |
| Quantization |
float16 |
| Inference backend |
faster-whisper |
