xb2g/thai-whisper-small

Thai Whisper Small

Fine-tuned Whisper Small model for Thai Automatic Speech Recognition.

Training Details

• Base Model: openai/whisper-small (244M parameters)
• Dataset: Thai Elderly Speech (19,200 samples, 17 hours)
• Epochs: 3
• Loss Reduction: 86% (1.93 → 0.27)

Training History

Epoch	Train Loss	Test Loss	Improvement
1	1.9345	0.2274	Baseline
2	0.2893	0.2328	-85%
3	0.2691	0.2217	-86%

Usage

import torch
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load processor from base model
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

# Load base model
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Load fine-tuned weights
model.load_state_dict(torch.load("model.pt", map_location="cpu"))
model.eval()

License

MIT License