Articulatory-based speaker recognition

Abstract

This paper presents a new methodology for computational speaker recognition based on a mathematical model of articulatory speech production. The method, based on articulatory phonology is tested on the MOCHA database for recognizing a male speaker and a female speaker. From an engineering perspective, in articulatory phonology one is interested in the trajectories over time of a set of articulators. These time trajectories are associated with the production of speech. The basic phonological unit in articulatory phonology is the articulatory gesture, which is defined as a dynamic system specified by a characteristic set of parameters. This dynamic system receives as inputs a target state and a set of parameters that tune the system to the desired action. The output is the solution of the state equation, i.e, the state trajectory, where the state is formed by the x-y positions of the important articulators that describe human speech. The state trajectory is then mapped to the output speech waveform by emulating the human vocal tract, through the observation equation and the MFCCs frequency description. A simplification of this model will be used in this paper for speaker recognition with 100 percent success in recognizing a male and a female speaker. © 2013 Acoustical Society of America.