ADDING A PARAMETRIC APPROACH TO FORENSIC SPEAKER RECOGNITION
Franck Marescal · 2001
In the field of speaker recognition in forensic science, experts attempt to value the confidence of their results. While ideal binary decisions are unreachable, some numerical methods permit quantifying such confidence. This can be achieved using signal processing methods combined with a Bayesian approach to probabili- ties. Both recorded samples to be compared must be independently analysed in order to estimate the quality of available data. The continuation of the procedure depends on the reliability of these results. A quantitative step is then performed to compute models of speakers' voice using spectral and prosodic parameters. Bayesian theory on probabilities is implemented to interpret the results. It gives a fundamental rela- tion between prior and posterior odds of recognition through the likelihood ratio. This ratio helps the court in the final decision of a judgement because it gives the de- gree of importance of the evidence.