Data strategies in forensic automatic speaker comparison

Abstract

Automatic speaker recognition (ASR) is a method used in forensic speaker comparison (FSC) casework. It needs collections of audio data that are representative of the case audio in order to perform reference normalization and to train a score-to-LR function. Audio from a certain minimum number of speakers is needed for each of those purposes to obtain relatively stable performance of ASR. Although it is not possible to set a hard cut-off, for the purpose of this work this number was chosen to be 30 for each, and 60 for both. Lack of representative data from that many speakers and uncertainty about what exactly constitutes representative data are major reasons for not employing ASR in FSC.An experiment was carried out in which a situation was simulated where a practitioner has only 30 speakers available. Several data strategies are tried out to handle the lack of data: leaving out reference normalization, splitting the 30 speakers into two groups of 15 (ignoring the minimum of 30) and a leave 1 or 2 out strategy where all 30 speakers are used for both reference normalization and calibration. They are compared to the baseline situation where the practitioner does have the required 60 speakers. The leave 1 or 2 out strategy with 30 speakers performs on par with baseline, and extension of that strategy to the full 60 speakers even outperforms baseline. This shows that a strategy that halves the data need is viable, lessening the data requirements for ASR in FSC and making the use of ASR possible in more cases.