Recognizing a voice from its model

The authors are investigating a potential solution to the `large population' speaker identification problem by characterizing the voice by the entailments in two different kinds of models. These entailments are found in the representational models of Neuro-Linguistic Programming (NLP) and in the model of the mechanics of the voice as revealed by the continuous wavelet transform (CWT). Results to date have been obtained from examining samples in the TIMIT database and human subjects. Local features correlated with individual speakers for selected vowel sounds have been found in CWT space. Features of NLP representation systems have also been found and are with voice features for speakers whose NLP representation systems are known a priori. Gaussian mixture models are used to calculate probability density functions from the local feature distributions. This speaker identification strategy combines three elements of novelty. First, it exploits the fact that the two-dimensional CWT of a one-dimensional signal can be interpreted as an image, and uses feature extraction techniques first developed for image processing. Second, this is the first known study in which voice waveforms are systematically studied to identify features that are attributed to the speaker's mental representation. Third, the reliability of the identification will be strengthened by combining entailments from these two completely different aspects of the speaker's identity, the mechanical aspects of the speaker's vocal tract, and the pattern of representation.