Intervalo de evidência e pareamento fuzzy utilizando relação sinal ruído aplicados à comparação forense de locutores

Abstract

Forensic Speaker Comparison (FSC) is an analysis of two voice samples to infer the similarity of their features. One voice sample is a trace of a criminal fact, while the other is from a known individual. FSC differs from voice biometrics in several ways. Biometrics allows the control of some variables not controllable in FSC, among them, the recording device, the channel noise, the quantity and duration of the samples, and speaker cooperation. Also, there is no risk of associating (or failing to associate) an innocent to guilty part. The new paradigm shift of forensic sciences can be characterized as quantitative databased implementation of the likelihood-ratio framework with quantitative evaluation of the reliability of results. This thesis presents two lines of experimental work within the new paradigm shift. The experiments were conducted using two databases for training and validation. Parameters such as noise type and intensity of speech contamination, as well as the duration of the speech sample were evaluated. The first line developed a synthetic solution for the Full Bayesian Significance Test (FBST) over the mean with unknown variance, proposing an interval estimation, hereinafter referred to as evidence interval, applicable to the FSC. In the experiments, variation of the SNR showed the evidence interval reduced error rates (Type I and Type II) by approximately 6.4%, surpassing other evaluated methods. The second line of investigation proposed the use of Spectrographic Signal-to-Noise Ratio (S2NR) measures to separate the signal into fuzzy sets for the comparison of speakers. Moreover, it was also proposed an adaptation for the calculation of the Baum-Welch statistics. Experiments with combinations of SNR and speech size showed the proposed method reduced false-positive rates by 35.7%, rising also accuracy in 4.5%, compared to other evaluated techniques. The results are promising and are being used experimentally at the Instituto de Criminalística of Polícia Civil de Minas Gerais.