Noise Performance of Various Speaker Verification Algorithms
Peter J. Kootsookos · 1992
This report details a comparison between three speaker verification algorithms used in the presence of noise. The algorithms compared are dynamic time-warping, vector quantization and a recurrent neural network approach. In an attempt to cope with the noise, the speech was enhanced using Kalman smoothing. Improvements in the speed of the Kalman filtering algorithm (over the previously reported algorithm) by reducing the order of the linear predictive model use and the resulting performance degradation are also reported here. We indicate where various further improvements in the system may be made.