Early Results for a Non-parametric Hidden Markov Model for Talker Characterization

This paper defines talker characterization problem and contrasts it to the related problems of talker verification and identification. The requirement is to automatically separate and tag talkers in a conference environment using a microphone array and unconstrained speech. An approach which uses a continuous ergodic HMM organized over broad phonemic categories using non-parametric observation probabilities is presented. Talkers are characterized by forming characterization vectors, where a characterization vector is a collection of likelihoods from a referent set of models. In the first investigation, the experiments use close-talking speech from the TIMIT database, rather than from microphone-array speech, in order to evaluate the characterization method for the more ideal setting.