Controle de acesso a ambiente restrito, a partir da identidade vocal, utilizando coeficientes MFCC e classificador K-means.

Abstract

The machine learning area is a great ally to ensure privacy and security, as it promotes advances in the methods used for access control. The use of techniques for Automatic Recognition of the Voice Identity of Speakers, for authentication purposes, represents one of these advances. Given the above, this article aims to present a system for automatic verification of the vocal identity of speakers, seeking to apply it for authentication and release of access to a restricted environment. The system is based on a pattern recognition task, divided into two stages: training and verification. In the training, techniques were applied for pre-processing the signal (pre-emphasis, division into frames and windowing), extraction of characteristics (Mel-Frequency Cepstral Coefficients - MFCC) and construction of a representative pattern of the vocal identity of each speaker (clustering). In the verification, pre-processing of the signal, extraction of characteristics and authentication occurred, the latter based on the comparison between the test characteristics and the previously stored pattern of the speaker. In the logic decision, thresholds were used for the authentication of an announcer (acceptance, rejection and indeterminacy). The results obtained demonstrate a correct authentication of the speaker in 81% of the cases and a rate of 94.89% of rejection of imposters, proving the efficiency of the proposed approach.