Feature extraction for robust speech recognition based on maximizing the sharpness of the power distribution and on power floori