Whisper-Small-ASR ( Burmese )
This model is a fine-tuned version of openai/whisper-small for Automatic Speech Recognition (ASR) in Burmese .
Model Overview
- Architecture: Whisper Small
- Parameters: ~244M
- Task: Automatic Speech Recognition (ASR)
- Language: Burmese (
my) - Sampling Rate: 16,000 Hz (Mono)
Dataset Details
The model was trained on a standardized, high-quality Burmese speech corpus:
- Total Duration: ~22 Hours of audio
- Total Utterances: 24,560 WAV files (16 kHz, 16-bit PCM, Mono)
- Total Speakers: 13 Synthetic Burmese Speakers
- Speakers: ပီယ ၊ ဝါစာ ၊ သီရိ ၊ ဒီပ ၊ အက္ခရာ ၊ သဒ္ဒါ ၊ သရ ၊ တာရာ ၊ ကဝိ ၊ သုတ ၊ ပသာဒ ၊ ဂီတ ၊ နန္ဒ
- Training & Validation: 11 speakers (20,699 train utterances, 2,300 validation utterances)
- Test Set (Held-Out): 2 unseen speakers ( ဂီတ & နန္ဒ , 1,561 utterances )
- Text Preprocessing: Standardized Myanmar Unicode (
\u1000-\u109F) with unified word segmentation and punctuation removal. The held-out test split evaluates zero-shot acoustic generalization across completely unseen voices.
Training Configuration
| Parameter | Value |
|---|---|
| Effective Batch Size | 64 (8 per device x 4 gradient accumulation x 2 GPUs) |
| Peak Learning Rate | 2.5e-5 |
| Learning Rate Scheduler | Cosine |
| Warmup Steps | 200 |
| Total Steps | 2,000 (~6 Epochs) |
| Mixed Precision | FP16 |
| Gradient Checkpointing | Enabled |
| Augmentation | None |
Evaluation Results
| Stage / Evaluation | Step | Train Loss | Val Loss | WER (%) | CER (%) | SER (%) | DER (%) | IER (%) | chrF |
|---|---|---|---|---|---|---|---|---|---|
| Baseline (Untrained) | 0 | - | - | 610.77 | 441.11 | 100.00 | 44.62 | 510.77 | 0.00 |
| Validation Set | 250 | 0.9354 | 0.1031 | 60.56 | 14.25 | 97.70 | 18.50 | 1.39 | 74.65 |
| Validation Set | 500 | 0.3638 | 0.0493 | 33.98 | 6.46 | 82.16 | 4.85 | 4.22 | 88.13 |
| Validation Set | 750 | 0.1874 | 0.0387 | 28.06 | 5.14 | 71.61 | 4.94 | 3.03 | 91.09 |
| Validation Set | 1000 | 0.0869 | 0.0338 | 23.82 | 3.90 | 65.54 | 3.88 | 3.27 | 93.54 |
| Validation Set | 1250 | 0.0842 | 0.0314 | 23.44 | 3.70 | 64.97 | 5.46 | 2.16 | 94.00 |
| Validation Set | 1500 | 0.0319 | 0.0339 | 21.29 | 3.37 | 60.33 | 3.31 | 3.47 | 94.58 |
| Validation Set | 1750 | 0.0080 | 0.0372 | 20.96 | 3.25 | 58.98 | 3.87 | 2.91 | 94.80 |
| Validation Set | 2000 | 0.0043 | 0.0375 | 20.70 | 3.21 | 58.59 | 3.87 | 2.78 | 94.86 |
| UNSEEN TEST (Final) | Final | - | - | 28.22 | 5.74 | 84.38 | 6.34 | 2.47 | 89.93 |
Usage
import torch
from transformers import pipeline
pipe = pipeline(
'automatic-speech-recognition',
model='thantzinphyo/Whisper-Small-ASR',
device='cuda:0' if torch.cuda.is_available() else 'cpu',
)
output = pipe('audio.wav', generate_kwargs={"language": "my", "task": "transcribe"})
print(output['text']) # ငါ့ အချစ်ရေးကျတော့ ဘာလို့ မကောင်းတာလဲ ဆိုတာ ဘယ်သူမှ မပြောနိုင်ဘူး
License
Apache-2.0