The only advice I can give is this: Train the model with more voice data, or train the model on AITCH.