Pole-zero modelling of speech for use in nasality based speaker recognition.

Abstract: In this bachelor thesis we performed a comparison between different methods of fitting pole-zero filters to data, and their usefulness for nasality-based speaker recognition in particular. It is believed that nasality has low intra-speaker variation and high inter-speaker variation, and that polezero filters are good at capturing the nasal characteristics. We describe to the extent possible, the theory underpinning the various methods, and then compare them in various ways. We used simulated speech data, to see to what extent the methods provided good estimates of the true filters, when noise had been introduced afterwards. Another way we compared the methods was to determine how well they perform with respect to classifying real speech data, when various features had been extracted with help from the computed filters. Our results were that the all-pole method was superior to the other methods that we considered, at speaker recognition on nasal phonemes, contradicting our hypothesis.