Detecting an imposter in telephone speech

This paper presents initial results on imposter detection in telephone speech. The imposter detector problem is defined in terms of a real-world security problem. Perceptual studies are then presented. These studies present a good estimate on the difficulty of the task at hand; it is found that humans classify approximately 85.6% of our benchmark utterances correctly. To design an automatic imposter detector, features which elicit speaker differences are studied. A baseline system based only on 20th order Linear Predictive Coefficients (LPC) classifies 75.0% of the test set correctly. By extracting features only in vowel and semi-vowel regions, i.e. where the all-pole model of the linear predictor is most accurate, the classification performance is increased to 80.0%. Further features such as average energy and median pitch result in a correct classification rate of 83.7% comparable to the perceptual benchmarks