# 5-gram language model optimization

Vietnamese automatic speech recognition model based on wav2vec2 architecture, pre-trained with 13,000 hours of unlabeled YouTube audio and fine-tuned on 250 hours of labeled data

A Romanian speech recognition model fine-tuned based on facebook/wav2vec2-xls-r-300m, trained on Common Voice 8.0 and Romanian speech synthesis datasets, ranked first in Romanian speech recognition in the HuggingFace Robust Speech Challenge.