A spectrum smoothing method for speaker verification

In speech processing, speech signal is usually processed frame by frame due to the non-stationary characteristic of speech. In this paper, a frequency-domain averaging based frame smoothing method is proposed. Besides the conventional frame shift, we introduce a short time shift to create several frames around current frame. Then we take the average of power spectrum for these frames. The average will be treated as a new frame instead of current frame. The new frame is considered to retain more integrated phonetic information than conventional frames. An experiment on speaker verification task showed that this method could improve the performance of speaker verification. The evaluation tasks performed on the NIST SRE 2008 database showed that our proposed method could achieve a better verification performance when compared with the conventional framing methods.