Measuring surprisal in sound sequences

Sensory input that violates prior expectations attracts attention, making unpredictability an important perceptual property to measure. In the auditory modality, knowing what sounds will be perceived as surprising, and therefore salient, is relevant both for studying vocal communication and for applied purposes such as managing noise pollution. Focusing on sequences of animal vocalizations and environmental sounds as ecologically important acoustic stimuli, I describe and benchmark several algorithms for measuring their perceived unpredictability. Information-theoretical approaches include Shannon surprisal and Bayesian surprise, both implemented here to detect deviant stimuli based on distributional acoustic properties. The second group of algorithms is based on detecting spectro-temporal recurrence assessed with autocorrelation functions (ACF surprisal) and self-similarity matrices (SSM novelty). The third approach uses neural networks. Based on the ratings of the predictability of 300 synthetic acoustic sequences by 195 human listeners, Shannon surprisal and SSM novelty capture the perceived unpredictability that is due to spectral variability, whereas ACF surprisal taps into the perceptual impact of irregular rhythm. Most algorithms converge on the time scale of about 1 s as the most perceptually relevant for spectral variability, which is consistent with the hypothesis that the perception of unpredictability stems from a relatively limited amount of auditory input held in short-term memory. Together, the presented open-source algorithms offer powerful and flexible tools for measuring acoustic surprisal and studying auditory attention, while the corpus of predictability ratings offers a resource for future benchmarking. All code and data are freely available from the R package soundgen and supplementary materials at https://osf.io/bgzvc .