A new study has offered a detailed computational map of what makes a sound sequence feel predictable—or unexpectedly startling. Published in Behavior Research Methods, the research compares several algorithms designed to measure auditory “surprisal,” a technical term for the degree to which incoming sensory information violates expectations. The work focuses on sequences of animal vocalizations and environmental sounds, but its implications extend to music, speech, urban noise, alarms, wildlife monitoring and the broader science of attention. By making the analysis tools freely available, the study could give researchers and sound designers a practical way to quantify why certain noises seem to demand immediate notice.
The central idea is simple: the brain is constantly forecasting what will happen next. When the next sound differs from that prediction, processing intensifies and attention may be drawn toward it. A dog bark arriving earlier than expected, a repeated call suddenly changing pitch, or an alarm pattern breaking its rhythm can all produce a sense of surprise. Yet unpredictability is not a single acoustic property. A sequence may be irregular because its timing changes, because its frequency and timbre fluctuate, or because familiar patterns disappear altogether. The new research therefore tests several methods that capture different kinds of acoustic change rather than treating surprise as one universal number.
To benchmark the algorithms, researcher Andrey Anikin created 300 synthetic sound sequences, each approximately seven seconds long, using the open-source R package soundgen. The recordings were based on 25 realistic prototypes, including animal calls, human nonverbal vocalizations and environmental noises. Twelve versions of each prototype were generated by changing two parameters. “Randomness” altered the timing and irregularity of pauses between sounds, while “temperature” increased the spectro-temporal variability of the individual calls or syllables. Every sequence was normalized to the same peak amplitude, ensuring that differences in perceived unpredictability were not simply caused by some recordings being louder than others.
The sequences were rated online by 195 adults living in the United Kingdom. Each participant listened to a randomized selection and judged how predictable each sequence was on a scale from “not at all predictable” to “completely predictable.” A multilevel Bayesian model was then used to estimate population-level ratings while accounting for differences among listeners, sound prototypes and individual recordings. The resulting estimates showed a continuous range of perceived predictability. In general, variation in the individual sounds had a stronger relationship with unpredictability judgments than irregular rhythm did. Variability in syllable duration was especially important, followed by variation in the intervals between sounds and changes in spectral features such as peak frequency and spectral centroid.
One of the most direct approaches tested was Shannon surprisal, calculated from the negative logarithm of the probability assigned to an observation. In information theory, an event with probability p has surprisal equal to −log₂(p), meaning that rare events carry more information. For raw acoustic signals, however, the probability of every possible sound cannot be calculated directly. The study therefore used a practical approximation: the current energy in each frequency band was compared with the mean and variance of recent energy values. A Gaussian probability density estimated how well the new observation fit the recent acoustic distribution. A close match produced low surprisal, while a large deviation produced a high value.
This distribution-based method performed surprisingly well. When applied to mel-frequency spectrograms, Shannon surprisal correlated with human judgments at approximately r > .6 under optimized settings. Its strongest performance generally came from an analysis window lasting roughly 700 to 1,000 milliseconds. That result suggests that the recent past most relevant to auditory predictability may be close to one second long. The finding is consistent with research on echoic memory, the short-lived auditory store that preserves enough recent sound information for the brain to detect changes and compare incoming signals with what has just been heard.
The study also examined Bayesian surprise, which differs from Shannon surprisal by measuring how much an observation changes an entire probability distribution. The method was implemented using the Kullback–Leibler divergence between two Gaussian distributions representing acoustic properties before and after a new observation. In principle, this approach can distinguish between a rare event that confirms an existing model and one that forces the listener to revise that model. In practice, Bayesian log-surprise was less closely related to the listeners’ overall ratings, reaching a best correlation of approximately r = .25. Although its moment-to-moment contour often closely resembled Shannon surprisal, the two methods did not summarize perceived sequence-level unpredictability in the same way.
A second family of algorithms focused on repetition rather than probability distributions. The new autocorrelation-based measure, called ACF surprisal, estimates how strongly a sound sequence repeats at different time lags. For example, a series of barks arriving at nearly regular intervals produces a peak in the autocorrelation function corresponding to the typical time between barks. If the next bark arrives late, early or with a different acoustic structure, the peak weakens and the algorithm registers positive surprisal. If the new bark reinforces the established pattern, the peak becomes stronger, producing negative surprisal. This makes ACF surprisal bipolar: positive values indicate a disruption, while negative values indicate reinforcement of a recurring pattern.
ACF surprisal was able to capture both rhythm and spectral deviations, but its performance depended strongly on how the signal was analyzed. Longer windows of approximately four seconds were useful for measuring rhythmic regularity, often with only the amplitude envelope required. Shorter windows near one second, combined with several frequency bands, were better for detecting spectral variability. The mean ACF measure showed only a modest association with the ratings, around r = .2, although its negative component—representing reinforced recurrence—performed better, reaching roughly r = .4. The author notes that positive spikes may still be valuable for detecting brief local deviations, even if they do not predict how unpredictable listeners find an entire sequence.
The strongest recurrence-based results came from self-similarity matrix novelty, a technique originally developed for analyzing musical structure. In this approach, short acoustic segments are converted into feature vectors and compared with other segments throughout the recording. The resulting matrix shows where similar sounds recur. A checkerboard-shaped kernel then compares neighboring regions: novelty is high when each region is internally consistent but the two regions differ from one another. Unlike the other measures, this calculation uses information from both sides of a time point, making it unsuitable as a strictly real-time detector without introducing a delay. With kernels lasting about 900 to 1,800 milliseconds, SSM novelty closely tracked human judgments, particularly when spectral variation was responsible for the perceived unpredictability.
The researchers also tested a neural-network system called PAAudioIC, originally designed to estimate surprisal in music. Using its default settings, the model achieved a correlation of approximately r = .40 with the human ratings. Its output was closely associated with both Shannon surprisal and the temperature parameter used to generate the stimuli, suggesting that it primarily captured short-term spectral change. Rather than declaring one algorithm the winner, the study shows that the measures are complementary. A combined index incorporating Shannon surprisal, SSM novelty and long-window ACF surprisal reached r = .72 with human ratings, while adding variability in syllable duration and interburst intervals raised the correlation to approximately r = .78.
The findings do not establish a universal formula for auditory surprise. The benchmark consisted of relatively clean, synthetic sequences in which individual sounds were separated by silence, and the optimized settings were chosen using the same dataset on which performance was assessed. That creates a risk of overfitting, and natural soundscapes, overlapping conversations, music and speech may require different models. The measures also estimate acoustic unpredictability, not attention itself: whether a sound captures attention depends on goals, context, emotional meaning, source separation and prior experience. Even so, the study provides an unusually broad toolkit for investigating how the brain detects change. Its open-source code, data and sound-analysis functions could help researchers compare methods across species, environments and perceptual tasks—and may eventually help explain why some sounds fade into the background while others instantly become impossible to ignore.
Subject of Research: Auditory surprisal and the perceived predictability of sound sequences
Article Title: Measuring surprisal in sound sequences
Article References: Anikin, A. (2026). “Measuring surprisal in sound sequences.” Behavior Research Methods, 58, Article 278.
Image Credits: AI Generated
DOI: https://doi.org/10.3758/s13428-026-03153-3
Keywords: Auditory attention, auditory surprisal, predictability, Shannon surprisal, Bayesian surprise, autocorrelation, self-similarity matrix, acoustic novelty, sound sequences, psychoacoustics
