Abstract
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.