Whisper Tiny Myanmar — ShopVoice ASR
A fine-tuned OpenAI Whisper Tiny model for Myanmar-language ShopVoice applications.
Based on chuuhtetnaing/whisper-tiny-myanmar and trained with the author’s Myanmar ShopVoice data.
Dataset
Dataset used to train: thantzinphyo/myanmar-shopvoice
The publicly available dataset is provided for demonstration purposes; however, the definitive model was trained using a moderately sized ShopVoice dataset compiled by the authors
Currently, the model has been trained on 237 distinct dish categories sourced from traditional Myanmar eateries.
Training Results
| Step | Eval Loss | WER | CER | Quantity Acc. | Exact Sentence Acc. | Product Name Acc. | Catalog Top-1 Acc. |
|---|---|---|---|---|---|---|---|
| 100 | 0.3237 | 48.18% | 14.34% | 86.13% | 23.36% | 27.01% | 29.20% |
| 200 | 0.0509 | 31.39% | 8.77% | 94.16% | 46.72% | 48.54% | 51.82% |
| 300 | 0.0360 | 25.73% | 7.10% | 96.72% | 55.84% | 57.30% | 60.58% |
| 400 | 0.0311 | 24.27% | 6.25% | 96.35% | 58.76% | 60.95% | 64.23% |
| 500 | 0.0289 | 22.99% | 6.00% | 97.08% | 60.58% | 62.04% | 64.23% |
| 600 | 0.0279 | 20.62% | 5.49% | 97.45% | 63.50% | 64.96% | 67.88% |
| 700 | 0.0272 | 19.34% | 5.18% | 97.45% | 65.69% | 66.79% | 69.34% |
| 800 | 0.0272 | 19.53% | 5.39% | 97.45% | 65.33% | 66.42% | 69.34% |
| 900 | 0.0272 | 19.53% | 5.30% | 97.08% | 65.33% | 66.79% | 69.71% |
| 1000 | 0.0272 | 19.34% | 5.18% | 97.45% | 65.69% | 66.79% | 69.71% |
Usage
from transformers import pipeline
pipe = pipeline(
"automatic-speech-recognition",
model="thantzinphyo/whisper-tiny-myanmar-shopvoice",
)
result = pipe(
"audio.wav",
generate_kwargs={"language": "my", "task": "transcribe"},
)
print(result["text"])
For best results, use 16 kHz mono audio.
Attribution
- Model: thantzinphyo/whisper-tiny-myanmar-shopvoice
- Base model: chuuhtetnaing/whisper-tiny-myanmar
- Dataset: thantzinphyo/myanmar-shopvoice
- Author: Thant Zin Phyo