Assessing phoneme distribution for speech modeling