ISCA Archive SPAC 1992
ISCA Archive SPAC 1992

A comparison of feature performance under degraded speech in speaker recognition

J. P. Openshaw, Z. P. Sun, J. S. Mason

This paper assesses several recently proposed strategies for combatting the serious adverse effects of additive noise and spectral tilt. Direct comparisons of mel, PLP and their dynamic forms, RASTA extensions, norm and angle distortion measures, and feature combination via linear discriminant analysis (LDA) are made. For tilt we show that regression features (Δ) show only small degradation, although their performance generally is worse than that of their static counterparts. PLP is sensitive to tilt, but RASTA processing gives significant improvements. For noise we find effective robustness only when features from noisy data are incorporated into the model. Furthermore we show feature combination of mel and PLP-RASTA to outperform the more standard mel plus Δmel pairing.


Cite as: Openshaw, J.P., Sun, Z.P., Mason, J.S. (1992) A comparison of feature performance under degraded speech in speaker recognition. Proc. ETRW on Speech Processing in Adverse Conditions, 119-122

@inproceedings{openshaw92_spac,
  author={J. P. Openshaw and Z. P. Sun and J. S. Mason},
  title={{A comparison of feature performance under degraded speech in speaker recognition}},
  year=1992,
  booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
  pages={119--122}
}