An important task in criminal justice is to evaluate the accuracy of eyewitness testimony. In this study, we examined if machine learning could be used to detect accuracy. Specifically, we examined if support vector machines (SVMs) could accurately classify testimony statements as correct or incorrect based purely on the nonverbal aspects of the voice. We analyzed 3,337 statements (76.61% accurate) from 51 eyewitness testimonies along 94 acoustic variables. We also examined the relative importance of each of the acoustic variables, using Lasso regression. Results showed that the machine learning algorithms were able to predict accuracy between 20 and 40% above chance level (AUC = 0.50). The most important predictors included acoustic variables related to the amplitude (loudness) of speech and the duration of pauses, with higher amplitude predicting correct recall and longer pauses predicting incorrect recall. Taken together, we find that machine learning methods are capable of predicting whether eyewitness testimonies are correct or incorrect with above-chance accuracy and comparable to human performance, but without detrimental human biases. This offers a proof-of-concept for machine learning in evaluations of eyewitness accuracy, and opens up new avenues of research that we hope might improve social justice.

Criminal-law workers must evaluate the accuracy of eyewitness testimony, which includes nonverbal speech characteristics that may carry clues. Using a large collection of testimony statements (N = 3344), Study 1 showed that eyewitnesses use a higher speech rate, fewer pauses, and greater acoustic amplitude when remembering correctly and more confident in their answers. In Study 2, observers with vs. without language comprehension (Swedish vs. American and Indian samples) judged the accuracy of testimony statements and judged the eyewitness’ confidence. Participants performed at above-chance levels regardless of nation or language comprehension, suggesting universal cues to accuracy. Participants gave higher confidence ratings to correct statements, and this calibration was higher with greater cultural proximity to the testimony source. Observers’ confidence judgments were significantly correlated with amplitude, which suggests they used valid acoustical cues. Taken together, results demonstrate distinct vocal markers of eyewitness accuracy, which are detected by listeners from diverse cultures.

In two studies, we examined if correct and incorrect statements in eyewitness testimony differed in semantic content. Testimony statements were obtained from participants who watched staged crime films and were interviewed as eyewitnesses. We analyzed the latent semantic representations of these statements using LSA and BERT. Study 1 showed that the semantic space of correct statements differed from incorrect statements; correct statements were more closely related to a dominance semantic representation, whereas incorrect statements were more closely related to a communion semantic representation. Study 2 only partially replicated these findings, but a mega-analysis of the two datasets showed different semantic representations for correct and incorrect statements, with incorrect statements more closely related to representations of communion and abstractness. Given the critical role of eyewitness testimony in the legal context, and the generally low ability of fact-finders to estimate the accuracy of witness statements, our results strongly call for further research on semantic content in correct and incorrect testimony statements.

Legal practitioners sometimes obtain eyewitness testimonies from non-native language speakers, yet few studies examine the effects of language in investigative interviews. Here, we investigate how testifying in a non-native vs. native language affects memory accuracy, susceptibility to suggestions, and witnesses perceived credibility. After viewing a mock-crime film, participants in Study 1 (N = 121) testified through (1) free recall, (2) cued recall and (3) the Gudjonsson Suggestibility Scales adapted to the crime scenario used in the present study either in their native (Swedish) or a non-native (English) language. They also rated their confidence in their memories, self-perceived credibility and cognitive effort. Native and non-native witnesses did not differ in memory accuracy, susceptibility to suggestions, self-rated credibility or cognitive effort. Non-native (vs. native) speakers did however report lower confidence in their memories. In Study 2, another group of participants (N = 202) were presented with the testimonies from Study 1, and judged witnesses' credibility. Non-native witnesses were judged as less credible than native speakers. Thus, while the lower confidence exhibited by non-native eyewitnesses did not correspond to their actual memory accuracy, it influenced observers' judgments of their performance. The results provide important knowledge for legal practices when evaluating the reliability of testimonies from non-native vs. native speaking eyewitnesses.

In two studies, we examined vocal characteristics of accuracy. Participants watched a staged-crime film and were interviewed as eyewitnesses. A mega- analysis showed that correct responses were uttered with 1) a higher pitch, 2) greater energy in the first formant region, 3) higher speech rate and 4) shorter pauses.

In two studies, we examined if correct and incorrect testimony statements were produced with vocally distinct characteristics. Participants watched a staged crime film and were interviewed as eyewitnesses. Witness responses were recorded and then analysed along 16 vocal dimensions. Results from Study 1 showed six vocal characteristics of accuracy, which included dimensions of frequency, energy, spectral balance and temporality. Study 2 attempted to replicate Study 1, and also examined effects of emotion on the vocal characteristic-accuracy relationship. Although the results from Study 1 were not directly replicated in Study 2, a mega-analysis of the two datasets showed four distinct vocal characteristics of accuracy; correct responses were uttered with a higher pitch (F0 [M]), greater energy in the first formant region (F1 [amp]), higher speech rate (VoicedSegPerSec) and shorter pauses (UnvoicedSegM). Taken together, this study advances previous knowledge by showing that accuracy is not only indicated by what we say, but also by how we say it.

An important task for the law enforcement is to assess the accuracy of eyewitness testimonies. Recent research show that indicators of effortful memory retrieval, such as pausing and hedging (e.g. “I think”, “maybe”), are more common in incorrect recall. However, a limitation in these studies is that participants are interviewed shortly after witnessing an event, as opposed to after greater retention intervals. We set out to mitigate this shortcoming by investigating the retrieval effort-accuracy relationship over time. In this study, participants watched a staged crime and were interviewed directly afterwards, and two weeks later. Half the participants also carried out a repetition task during the two-week retention interval. Results showed that the retrieval-effort cues Delays and Hedges predicted accuracy at both sessions, including after repetition. We also measured confidence, and found that confidence also predicted accuracy over time, although repetition led to increased confidence for incorrect memories. Moreover, retrieval-effort cues partially mediated between accuracy and confidence. 

Recent research has shown that incorrect statements in eyewitness testimonies contain more cues to effortful memory retrieval than correct statements. In two experiments, we attempted to improve judgments of testimony accuracy by informing participants about these effort cues. Participants read eyewitness testimony transcripts and judged statement accuracy. Performance was above chance in both experiments, but there was only a significant effect of the effort-cue instruction in Experiment 2. In Experiment 1, we also compared judgment accuracy between police detectives, police students and laypersons, and found no significant difference, in contrast to previous studies. Moreover, the current study corroborates previous findings that (a) judging testimony accuracy is a difficult task and (b) people spontaneously rely on effort cues to some extent when judging accuracy. However, a complete reliance on effort cues showed substantially better performance than relying on one's own judgments skills at best, and offered equal performance at worst.

Do sincere eyewitness testimonies contain objective markers of accuracy? We show that expressions of effort in memory retrieval predict eyewitness accuracy. Incorrect memories are recalled with greater effort than correct memories.