Robust text-independent speaker identification using bispectrum slice

Abstract

We propose to use a bispectrum slice for the Mel-frequency cepstrum coefficients as robust features, to be used in a Gaussian mixture model for text-independent speaker identification. In theory, higher order statistics can suppress additive Gaussian noise and save phase information, unlike autocorrelation based (power spectral) methods. Feature extraction is achieved through the Mel-frequency filter banks, the cosine transform and the logarithm operation to obtain cepstral coefficients. The performance of our proposed features is then compared with that of the classical Mel-frequency cepstrum coefficients under various noisy test utterances.