A study of LSF parameters for speaker recognition
Kuldip K. Paliwal · The Journal of the Acoustical Society of America · 1990
Linear prediction (LP) analysis of speech can lead to a number of parametric representations (such as cepstral coefficients, reflection coefficients, etc.), each of which provides equivalent information about the spectral envelope. The line spectral-pair frequency (LSF) representation is an alternate LP parametric representation. For speech coding and recognition applications, this representation has been recently found to be more useful than the other LP parametric representations. For speaker recognition application, the cepstral coefficient representation is traditionally used as an LP parametric representation. In this paper, the LSF representation for speaker identification is studied and its performance is compared with that of the cepstral coefficient representation. This representation results in speaker identificaton accuracy of 94% for the diagonal covariance matrix case and 95% for the full covariance matrix case; the corresponding results for the cepstral coefficient representation are 91.5% and 94.2%. Thus the LSF representation is better suited for speaker identification than the cepstral coefficient representation.