In this paper different techniques for pattern compression of a speaker verification system are presented. A bitrate reduction of the cepstral features to 8 bits by a better encoding shows only a small increase of the verification error. Trace segmentation and a simpler version without interpolation results in a given fixed length of the feature vector, but may be used only to reduce the length of longer phrases. Vector quantization especially in conjunction with DTW gives also the opportunity to use two pattern recognition techniques for different kind of speaker dependency within features.

Keywords: Speaker Verification, Cepstrum, DTW, Trace-Segmentation, VQ