SiangLao - Lao Voice-to-Text
Convert Lao speech to text with advanced AI models
Research Overview
- Models Compared: XLS-R-300M, XLSR-53, and HuBERT-Large
- Dataset: SiangLao/lao-asr-thesis-dataset
- Best Performance: XLS-R-300M achieved 12.5% Character Error Rate (CER)
- Training Method: 15 epochs with early stopping to prevent overfitting
Research Team
- Authors: Souphaxay Naovalath & Sounmy Chanthavong
- Advisor: Dr. Somsack Inthasone
- Institution: Computer Science Department, Faculty of Natural Sciences, National University of Laos
ASR Model
- Use clear and slow speech
- Record in a quiet environment
- Try different models to get the best results
- Audio quality affects transcription accuracy
SiangLao ASR System
Computer Science Department • National University of Laos