What This Study Shows: Tel Aviv University researchers demonstrate that AI systems can detect acoustic patterns in infant and animal vocalizations but fail to infer the meanings those signals convey to receivers. Using pre-linguistic toddler vocal sequences (distress, calling a caregiver, requesting food), the authors found that neural networks grouped sounds by acoustic similarity but missed contextual meanings and perceived urgency that humans detect. They recommend combining AI with behavioral tests, playback experiments, and neural data to model the listener's perceptual world and reliably decode communication.
AI Can Measure Calls — But Not Their Meaning: Israeli Study Using Babies Explains Why
German shepherd barking (illustrative) (photo credit: SHUTTERSTOCK)
Summary: A Tel Aviv University team shows that current AI systems reliably detect acoustic features of baby and animal vocalizations but fail to recover the meanings those signals carry for the receiver. The study argues for combining AI with behavioral and neural data to capture how listeners actually perceive and respond.
The dream of conversing with other species—from Dr. Dolittle to generations of pet owners—depends on more than detecting sound. A new paper in Current Biology, led by Prof. Yosef Yovel and colleagues at Tel Aviv University, demonstrates a fundamental limitation in common AI approaches to animal communication: algorithms reliably measure acoustic differences (volume, pitch, spectral shape), but they often miss the perceptual and contextual cues that determine meaning for a receiver.
How the Study Worked
The research team (Mor Taub, Inbal Arnon, Amiyaal Ilany, Mirjam Knörnschild, and Yoav Ram; supervised by Yovel) collaborated with groups at the Hebrew University of Jerusalem, the University of Edinburgh, the Natural History Museum—Leibniz Institute for Evolution and Biodiversity Science, and Humboldt University. To test whether AI can infer communicative meaning rather than acoustic structure, the authors used an unusually tractable dataset: vocal sequences from pre-linguistic human toddlers.
Tel Aviv University (credit: TEL AVIV UNIVERSITY)
Unlike most animal datasets, toddler vocalizations can be interpreted by adult listeners with reasonable confidence. The recordings included sequences produced in three contexts: distress, calling a specific caregiver (mother or father), and requesting food. The team analyzed these recordings using a classical acoustic pipeline and two deep neural networks — one trained on animal vocalizations and another trained on adult human speech — and asked the models to group the sounds by characteristics and (implicitly) by meaning.
Key Findings
Although the neural networks outperformed the classical acoustic method at detecting acoustic patterns, they failed to classify the toddler vocalizations by intended meaning. Models frequently merged acoustically similar sounds that served different communicative purposes and split acoustically dissimilar sounds that served the same purpose. Crucially, human listeners perceived ordered sequences that conveyed increasing urgency; none of the tested AI models recovered that ordering.
“Identifying acoustic patterns is not necessarily the same as deciphering meaning,” Yovel says. “To understand what an animal is 'saying,' we need to know how the animal receiving the message perceives it and responds to it.”
Why AI Falls Short
The authors argue that AI tends to optimize for the largest statistical differences in the data, which are not necessarily the differences that matter to a receiver's brain. The paper uses an illustrative example from stickleback fish: males respond primarily to red coloration during breeding, so a red-painted sphere elicits aggressive responses even if its shape is artificial. An algorithm scanning raw visual features would not automatically prioritize color unless guided to do so.
PROF. YOSEF YOVEL (credit: TAU)
Similarly, in vocal communication two sounds that look very different on a spectrogram can carry the same message to a listener, while acoustically similar sounds can mean different things depending on context and the listener's perceptual system.
Recommendations and Implications
The path toward truly deciphering animal (and pre-linguistic human) communication requires an integrated approach: AI for discovering patterns, plus behavioral observations, playback experiments, and — when possible — neural measurements from receivers. This combination helps identify which acoustic features receivers actually use and how context shapes interpretation.
The authors caution against overpromising what current AI can deliver. Pattern discovery is an important first step, but decoding meaning demands models and experiments designed around the receiver's perceptual world.
Conclusion
The study highlights a major conceptual shift: move beyond treating communication as a purely acoustic classification problem and instead model how signals are perceived and acted upon. Doing so will make AI a more reliable partner in efforts to understand animal minds and, ultimately, to build tools that can ask a cow about its health or better interpret the needs of pre-verbal infants.
Paper: "The Challenge of Decoding Animal Communication Using AI," Current Biology. Authors: Mor Taub, Inbal Arnon, Amiyaal Ilany, Mirjam Knörnschild, Yoav Ram; supervised by Yosef Yovel.
Help us improve.
