Likelihood ratio based authorship verification methods applied to forensic voice comparison tasks

Lexical and grammatical information such as the frequency of function words rarely contribute to forensic voice comparison conclusions. As such, forensic voice comparison is potentially missing out on the additional value that these types of features could bring to an analysis. In recognition of this, Sergidou et al. (2023) tested authorship analysis methods on speech data. They demonstrated that there is indeed speaker discriminatory power in word frequency information extrapolated from transcripts of speech recordings. The current work extends this line of research by testing further authorship verification methods that are considered state-of-the-art.

We used the transcribed data from 97 speakers of the WYRED corpus (Gold, 2020) participating in four different speaking tasks that are relevant to forensic speech casework. We applied three well-known methods in authorship verification - Cosine Delta, N-gram tracing, and the Impostors Method – to calculate likelihood ratios. The performance of the methods was assessed using the Cllr metric and this was below the threshold of 1 for the majority of the experiments. The best performing method was a variant of n-gram tracing that exploits both typicality and similarity information proposed in Nini (2023).

It is remarkable that applying authorship analysis methods to spoken data has only caught research attention very recently – this is likely a result of the distance that has traditionally existed between forensic linguistics and forensic speech science.

References

Gold, E. (2020). WYRED - West Yorkshire Regional English Database 2016-2019. [Data Collection]. Colchester, Essex: UK Data Service. 10.5255/UKDA-SN-854354
Nini, A. (2023). A Theory of Linguistic Individuality for Authorship Analysis (Elements in Forensic Linguistics). Cambridge, UK: Cambridge University Press.
Sergidou, E-K., Scheijen, N., Leegwater, J., Cambier-Langeveld, T, and Bosma, W. (2023). Frequent-words analysis for forensic speaker comparison. Speech Communication. 150. pp 1-8.