Optimizing the acoustic modeling from an unbalanced bi-lingual corpus