Can you try with a voice that was trained on that language? This research preview is more variable based on the voice chosen