法庭说话人识别新范式实证研究

Abstract

Objective: To demonstrate the core elements of the new paradigm of the evaluation of forensic evidence and its application in forensic practice through an empirical study on a real forensic voice comparison case. The new paradigm calculates a likelihood ratio value as strength of evidence statement using relevant data, quantitative measurement, and statistical models; and the performance of the system is empirically validated under conditions reflecting those of the case. Methods: In the case, the speaker of questioned identity on a telephone call was one of two known speakers. Five telephone recordings of each known speaker were made. Mel frequency cepstral coefficient (MFCC) values were extracted from each recording. Canonical linear discriminant functions were used to reduce dimensionality and for mismatch compensation. Likelihood ratios were calculated using univariate t distributions and regularized logistic regression separately. Results: Cross-validated testing on two procedures gave likelihood ratio values in the range of hundreds to tens of thousands for the former and in the range of hundreds for the later, always in the correct direction. The regularized logistic regression procedure took account of the small amount of training data and gave more conservative strength of evidence. Conclusion: The analysis reported in this paper demonstrates that the likelihood ratio framework can be used and the results empirically tested under the conditions of a real forensic voice comparison case.