Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels

Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels Daniel Frost1, Shunichi Ishihara2,3 1School of Language, Literature and Linguistics, The Australian National University, Australia 2Department of Linguistics, The Australian National University, Australia 3The Internet Commerce Security Laboratory, Federation University, Australia drbfrost@live.com.au, shunichi.ishihara@anu.edu.au Abstract This study is a pilot research that explores the effectiveness of a likelihood ratio (LR)-based forensic voice comparison (FVC) system built on non-native speech production. More spe- cifically, it looks at native Hong Kong Can- tonese-speaking male productions of English vowels, and the extent to which FVC can work on these speakers. 15 speakers partici- pated in the research, involving two non- contemporaneous recording sessions with six predetermined target words – “hello”, “bye”, “left”, “right”, “yes”, and “no”. Formant fre- quency values were measured from the trajec- tories of the vowels and surrounding segments. These trajectories were modelled using discrete cosine transforms for each for- mant (F1, F2 and F3), and the coefficient val- ues were used as feature vectors in the LR calculations. LRs were calculated using the multivariate-kernel-density method. The re- sults are reported along two metrics of per- formance, namely the log-likelihood-ratio cost and 95% credible intervals. The six best- performing word-specific outputs are present- ed and compared. We find that FVC can be built using L2 speech production, and the re- sults are comparable to similar systems built on native speech. 1 Introduction 1.1 Forensic voice comparison and the likeli- hood-ratio framework Forensic voice comparison (FVC) is the forensic science of comparing voices. It is most often used in legal contexts where the origin of voice samples is being debated. Typically, an FVC analysis in- volves the comparison of voice recordings of known origin (e.g. the suspect’s speech samples) with other voice recordings of disputed origin (e.g. the offender’s speech samples) (Rose, 2004). The FVC expert will apply statistical techniques on data extracted from speech sample evidence with the ultimate aim of assisting the trier of fact (e.g. judge(s)/jury) with their final decision. The trier of fact is faced with the task of making this decision by analysing the numerous probabilistic forms of evidence offered to them over the course of the trial. In fact, this decision is in itself a probabilistic statement, known as the posterior odds, and can be expressed mathematically as 1). 𝑝(𝐻|𝐸) 𝑝(𝐻̅|𝐸) (1) In 1), 𝑝(𝐻|𝐸) represents the probability of one hypothesis (e.g. the prosecution hypothesis – the suspect is guilty), given the various forms of evi- dence (e.g. DNA, fingerprint, voice, witness ac- counts etc.), and 𝑝(𝐻̅|𝐸) represents the probability of the alternative hypothesis (e.g. the defence hy- pothesis – the suspect is not guilty), given the evi- dence. In the context of FVC, 1) becomes: 𝑝(𝐻𝑆𝑆|𝐸) 𝑝(𝐻𝐷𝑆|𝐸) (2) In 2), 𝐻𝑆𝑆 represents the same-speaker hypothe- sis, and 𝐻𝐷𝑆 represents the different-speaker hy- pothesis. Before the trier of fact is able to make their decision of guilt or innocence, there may be more evidence that needs to …

Content truncated to 500 words