Data fusion for speaker parameterization by a possibility theory based method

摘要

In this paper, a speaker parameterization based on possibility theory has been developed in the experimental framework of speakers automatic identification from the acoustic data (MFCC coefficients) and anatomical data (length and thickness of the vocal cords). The data are modelled in the setting of the possibility theory which provides interesting tools of representing imprecision and uncertainty. Moreover, the constraints that govern this theory allow a wide choice for the combination of heterogeneous data. We are particularly interested by the adaptive combination rule proposed by Dubois and Prade. Thus, a fusion of acoustic and anatomical data in the form of possibility distributions is proposed. The resulting vector of this fusion is the vector representing the speaker who is the input of the second phase of the identification system that is the modeling phase.