Optimal acoustic model complexity selection in Polish medical speech recognition