Aplicação do método de fusão para verificação de locutor independente de texto

Abstract

This work presents an overview of text independent speaker verification, describing the basic operation of the system and the reviewing some important developments in speaker modeling and feature extraction from speech. Following, a point of improvement identified within the feature extraction stage leads to the main objective of this work: to determine one or more sets of coefficients relevant to speaker discrimination while minimizing the equal error rate (EER). The proposal is to replace the delta(Δ) and double-delta(Δ²) coefficients by a linear predictor code (LPC) for the mel frequency cepstral coefficients (MFCC). In addition, score level fusion is employed to combine the ouputs of MFCC-only and MFCC-LPC systems, as well as MFCC-only and MFCC-Δ-Δ² systems. In all cases, performance is evaluated with respect to variations of the signal to noise-ratio (SNR) in the tested audio. In addition, the work introduces a new Brazilian Portuguese speech repository containing free-speech from 155 males. Results and discussions are presented with a reflection on the expected outcomes, as well as general comments and observations. Finally, concludings remarks are made about the work, featuring future prospects regarding text independent speaker verification research. This work attained a 4% reduction in the EER compared to the reference system (MFCC-only), with best results occuring in the case fusion of MFCC-only and MFCC-Δ-Δ² scores.