To my surprise, when 'sr' is specified as the file sampling rate, the librosa is five times faster than the scipy. The scipy is always faster than the librosa.