Whisper-Tiny-Myanmar-UnFreezing

This model is a Full fine-tuned version of (https://huggingface.co/thantzinphyo/Whisper-Tiny-Myanmar-Partial-Freezing)) trained on thantzinphyo/burmese-speech-refined-openslr-80 Dataset.

Evaluation Results (Best Checkpoint: Epoch 7)

• Best Validation WER: 54.8060%
• Best Validation CER: 19.3867%
• Validation Loss: 0.0970

---

Full 30-Epoch Training Results

Training Loss	Epoch	Step	Validation Loss	WER (%)	CER (%)	Status
No log	1.0	9	0.3240	66.7549	25.2793
5.8635	2.0	18	0.2111	64.0653	23.5311
3.9579	3.0	27	0.1359	60.9347	21.8387
2.2134	4.0	36	0.1084	59.4797	21.7605
1.4742	5.0	45	0.0987	57.6720	20.7775
1.0824	6.0	54	0.0955	56.8783	20.4088
0.8153	7.0	63	0.0970	54.8060	19.3867	Best Checkpoint
0.5734	8.0	72	0.1063	55.9083	20.2301
0.3814	9.0	81	0.1170	55.4674	20.0290
0.2371	10.0	90	0.1294	56.4374	20.3083
0.2371	11.0	99	0.1398	56.7460	20.2692
0.1447	12.0	108	0.1463	56.6578	20.1408
0.1012	13.0	117	0.1492	56.7019	20.1631
0.0755	14.0	126	0.1564	56.7460	20.6881
0.0588	15.0	135	0.1650	56.1287	20.2804
0.0488	16.0	144	0.1697	57.2751	20.0402
0.0431	17.0	153	0.1701	56.4815	20.1017
0.0314	18.0	162	0.1740	55.0265	19.9788
0.0249	19.0	171	0.1784	56.3492	19.9397
0.0214	20.0	180	0.1768	54.8501	19.6995
0.0214	21.0	189	0.1824	56.3051	19.9453
0.0167	22.0	198	0.1866	55.6878	19.7107
0.0107	23.0	207	0.1900	55.5556	19.5766
0.0087	24.0	216	0.1929	55.3792	19.3476
0.0078	25.0	225	0.1961	56.5697	19.8168
0.0055	26.0	234	0.1962	56.1287	19.7554
0.0035	27.0	243	0.1982	56.3492	20.0235
0.0022	28.0	252	0.2009	57.3192	20.1296
0.0016	29.0	261	0.2019	57.0988	20.0011
0.0017	30.0	270	0.2022	56.8342	20.0681

---

Hyperparameter Configuration

This model was trained using Full End-to-End Fine-Tuning (All Layers Unfrozen) on top of thantzinphyo/Whisper-Tiny-Myanmar-Partial-Freezing

Hyperparameter	Value	Description / Rationale
Base Model	whisper-tiny-myanmar-phase1	Pre-aligned Burmese Whisper Tiny checkpoint
Training Mode	Full Unfreeze (End-to-End)	100% of parameters (37.7M) updated simultaneously
Learning Rate	5e-05 (5e-5)	Optimal sweet spot for fine-grained phonetic adaptation
LR Scheduler	Linear with Warmup	100 warmup steps followed by linear decay
Per-Device Batch Size	16	Optimized for GPU VRAM throughput
Gradient Accumulation	8	Simulates a stable global batch size
Effective Batch Size	128 (16 × 8)	Ensures low gradient variance and stable updates
Optimizer	AdamW	Weight decay of 0.01, $\beta_1 = 0.9$, $\beta_2 = 0.999$
Precision	FP16	Mixed-precision for accelerated training and memory efficiency
Epochs	30	Best checkpoint selected at Epoch 7 (load_best_model_at_end)

---