Forensic speaker recognition at the beginning of the twenty-first century - An overview and a demonstration

This paper has discussed some important aspects of forensic speaker recognition. It has emphasised that the task of a forensic speaker recognition expert is, after first quantifying the differences or similarities between the samples they are comparing, to estimate how much more likely this evidence is, assuming the samples have come from the same speaker than assuming they have not. The paper has, using Bayes' Theorem, explained why this is so, and shown how it is possible to do it, with a real example. It has also taken care to point out the shortcomings in the approach. There are shortcomings in the statistical modelling, which, although already highly sophisticated, is still not quite up to the complexities of speech, and there are also shortcomings in the availability of true reference populations. It should be clear from the paper that, properly done, FSR is a very complicated matter involving expert knowledge of, at least, linguistics, acoustics, statistics and signal-processing. It is not, as quite commonly supposed, a touchy-feely exercise whereby some individual gifted in recognising people by their voice listens to the recordings and makes their decision. It is also a painstaking, time-consuming, and, given the content of male telephone conversations in general, not very exciting undertaking. (The measurements for the er/fucken analysis demonstrated above took about ten hours. It took less than a second to press the key for the Likelihood Ratios, but a very long time to write the programs that derived them. The experiments to estimate the amount of reduction in LR for assumed correlated data took about three days.) Most important of all, however, the FSR expert needs to know how to interpret their findings forensically. This paper has shown how the Likelihood Ratio of Bayes' Theorem is now considered the proper construct for these findings - indeed, estimating the probabilities of the evidence under both prosecution and defence hypotheses must structure the whole forensic speaker recognition approach, as it should.