Reconhecimento automático de identidade vocal utilizando modelagem híbrida: paramétrica e estatística.

Abstract

This work deals with the application of a hybrid technique (parametric and statistical), which uses Linear Prediction Analysis, Vector Quantization, Neural Networks and Hidden Markov Models, to develop an automatic vocal identity recognition (identification) system, aiming at get alternatives to traditional algorithms. In order to achieve a faster and more robust system, and a pre-identification step is performed, followed by identification. The first step uses the fundamental frequency (F0) as the previous speaker separation parameter in general groups according to gender. The proposed method for estimating CF is efficient (99% correct classification), providing representative estimates of each speaker, thus reducing the number of speakers to participate in the later stage. The identification step uses Hidden Markov Models (HMMs) of Discrete Densities and Parametric Vector Quantization, with acoustic parameters obtained from Linear Prediction Analysis (LPC, Cepestral Weighted, Cepestral Delta and Weighted Cepestral Delta coefficients). Cepestral coefficients, followed by Cepestral Delta, provide higher identification rates. In terms of the vector quantizer dictionary design, three algorithms are evaluated: LBG (Linde-Buzo-Gray), KMVVT (Modified Kohonen with Neighborhood Centered Around Training Vector) and SSC (Competitive in Synaptic Space). The SSC algorithm is the most suitable for the dictionary design, leading to higher identification rates. Modeling by HMMs constitutes a "refinement" step of the identification process, being used when the distortion measurements obtained by comparing the speaker test pattern to be identified (acoustic characteristics vector) with the reference standards (dictionaries quantifier) ​​indicate "similarity" between vocal patterns. The technique applied in this work provides an automatic voice identity recognition system that has a high (97.8%) and significant average identification rate, low average rates of false acceptance (0.8%) and false rejection ( 1.5%) as well as high reliability (99.2%). The speaker identification system developed and, therefore, capable of efficiently discriminating the speakers from their vocal characteristics presenting, regardless of the speaker's gender, small intralocutor variations and large interlocutor variations.