A comparison of spectral smoothing methods for segment concatenation based speech synthesis