Application of automatic speaker verification techniques for forensic evidence evaluation

When applying speaker verification techniques for forensic evidence evaluation, many challenges arise. One such challenge is the interpretation of the verification results in terms of probability. The output of the existing speaker verification is a ratio between two likelihood values and hence does not have a very meaningful representation. In this paper, we introduce a method which outputs a posterior probability function, such that given the evidence, the probability it came from a certain suspect can be calculated. Another key challenge in forensic evidence evaluation is the practical difficulty in recreating the evidence environment in order to obtain the training voice samples from the suspect. For this we propose a novel approach based on the Baysean probability framework, which uses a pre-computed Universal Background Model (UBM). The speaker verification system is built around Mel Frequency Cepstral Coefficients (MFCC) for feature extraction and Gaussian Mixture Models (GMM) for speaker modeling. The optimum values for different parameters of these techniques are found experimentally. Our results demonstrate that a model with 32 mixtures and a dimension of size 13 gives the best performance, as a trade-off between the accuracy and the computational efficiency.