Whisper-Tiny-Myanmar-Partial-Freezing

Fine-tuned version of openai/whisper-tiny on SLR80: Crowdsourced High-Quality Burmese Speech Dataset ( https://www.openslr.org/80 ).

Best Model Selection Notice:

This repository contains the best checkpoint (Step 1,000), selected for its lowest CER of 19.60%.

---

How to Use (Python Inference)

import torch
from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="thantzinphyo/Whisper-Tiny-Myanmar-Partial-Freezing",
    device=0 if torch.cuda.is_available() else -1
)

result = pipe("your_burmese_audio.wav", generate_kwargs={"language": "my", "task": "transcribe"})
print("Transcribed Text:", result["text"]) # တချို့ ကလည်း ပြောတယ် မင်္ဂလာဆောင်တယ် ဆိုတာ အိမ်ထောင်ကျတာ ထောင်ကျတယ် ပေါ့

Complete Training & Evaluation History (Steps 0 to 1,500)

Evaluated on the isolated validation set (252 clips / ~0.42 hrs) and formatted in Percentage (%):

Step	Training Loss	Validation Loss	Character Error Rate (CER)	Word Error Rate (WER)	Checkpoint Status
0 (Baseline)	—	—	230.57%	213.09%	Pretrained Base (Hallucination)
100	17.2880	2.1209	78.09%	100.04%	Script Recognition
200	13.1279	1.6428	33.08%	81.22%	Phonetic Alignment
300	11.7845	1.6119	27.40%	73.99%	Syntax Formation
400	11.4952	1.6188	24.47%	69.96%	Sub-70% WER Barrier Broken
500	11.4360	1.6261	24.06%	70.84%	Vocabulary Consolidation
600	11.3406	1.6402	20.95%	66.37%	Particle Tuning
700	11.3297	1.6517	22.98%	69.09%	Regularization
800	11.3435	1.6443	21.23%	67.29%	Consonant Conjunct Refinement
900	11.3192	1.6535	20.37%	66.07%	Stable Convergence
1,000	11.2989	1.6744	19.60%	65.28%	Checkpoint (Published)
1,100	11.2955	1.6819	19.80%	64.97%	Lowest WER Record
1,200	11.2928	1.6945	19.71%	65.15%	High Precision
1,300	11.2914	1.7026	19.72%	65.19%	Cosine Learning Decay
1,400	11.2909	1.7060	19.94%	65.46%	Asymptotic Convergence
1,500	11.2908	1.7066	19.88%	65.37%	Final Step Completion

---

Initial Training Configuration

• Base Architecture: openai/whisper-tiny (37.8M parameters)
• Training Strategy: Phase 1 Partial Encoder Freeze — Layers 0–1 frozen, while Layers 2–3 and the Decoder were trainable (33.1M parameters / 87.66%)
• Effective Batch Size: 64 (16 per device × 4 gradient accumulation)
• Learning Rate: 1.5e-4 with a Cosine Scheduler and 150 warmup steps
• Dataset: 2,277 training clips and 252 validation clips using 100% OpenSLR-80 NFC-canonical transcriptions