Optimizing Data Selection for Automatic Speech Recognition in Low Resource Languages