Quadratic detectors for feature extraction in text-independent speaker authentication

Text-independent speaker authentication requires a feature set which is sensitive to characteristics of speakers while being reasonably invariant across utterances. Features obtained from cepstra-based techniques have long been used for speech recognition. However, these features are utterance dependent and sensitive to noise so that it may be more difficult to use them for robust speaker authentication. In this paper, a quadratic detector, which is closely related to quadratic time-frequency representations, is proposed to achieve the required utterance-invariant feature extraction. As demonstrated on data derived from the King Corpus, the features extracted using the quadratic detector can provide better classification accuracy than solely cepstral features.<<ETX>>