In a demonstration based on toddler voices, AI tools identified similarities between sounds, but not always the message conveyed. Researchers call for combining voice analysis with observation of listeners' reactions
Artificial intelligence can sort recordings of bats, whales, and birds based on their sound properties. But a group of similar-sounding voices doesn’t necessarily convey the same message. A paper by researchers at Tel Aviv University and colleagues, published in Current Biology, illustrates the gap between acoustic similarity and meaning using recordings of human toddlers.
The researchers didn’t try to talk to bats, and they didn’t show that any AI tool would always fail to decipher their voices. Their demonstration offers an important test of the field’s promise: To determine what a signal means, you also need to examine the context and the way the receiver perceives and responds to it.
Why sorting sounds is not enough
A computer system can measure frequency, intensity, duration, and repeating patterns in calls. It may easily identify two voices that sound similar, yet miss a situation where one is carrying a message of distress and the other is used to call for a parent. The opposite is also possible: two acoustically different voices can express a similar request.
The researchers' starting point is that meaning in communication is not a property of the sound wave alone. It also depends on the sensory world of the listening animal, the circumstances in which the signal is heard, and the response it evokes. Therefore, automatically grouping recordings according to their physical form is not equivalent to translating them.
The article is signed by Mor Taub, Prof. Inbal Arnon, Prof. Amiel Ilani, Dr. Miriam Kernschild, Dr. Yoav Ram, and Prof. Yossi Yuval. The researchers are affiliated with Tel Aviv University, the Hebrew University, and research institutions in Germany and the United Kingdom.
Toddlers as a test case
To test how similarity in sound relates to meaning, the team selected a small group of recordings from toddlers who had not yet acquired spoken language. The recordings were made in three contexts that could be identified by the situation in which they were produced: distress, addressing a mother or father, and requesting food. Unlike wild animal calls, here we can at least roughly know what was happening around the sound and how a person hearing it might interpret it.
The researchers compared a standard acoustic method with two neural networks: one trained on animal sounds and another trained on adult human speech. They looked at how the methods grouped the recordings, not just whether the computer found technical similarities between them.
The neural networks outperformed the standard acoustic method in the metrics tested, but the groups they created did not consistently match the contexts in which the sounds were made. Sometimes calls intended to convey a similar message were separated, and sometimes calls that carried different messages were combined. According to the university's announcement, the tools also did not capture the increase in urgency along a sequence of calls.
What can we learn from this about bats and whales?
The article is short and based on a small sample of toddler sounds. It does not prove that it will be impossible to decipher animal communication with artificial intelligence; it demonstrates why meaning cannot be identified based on acoustic similarity alone. In other species, the difficulty is even greater, because humans do not always know how the individual listening experiences the signal.
The researchers suggest combining the analysis of sound with behavioral observations, listening experiments in which signals are played to animals and their response is examined, and sometimes also measurements of nervous system activity. This way, it is possible to get closer to the question of what the signal does in the animal's world, beyond the question of how it appears on the computer screen.
Prof. Yuval, who studies, among other things, communication and vocal sensing in bats, emphasizes in the university's announcement that promises for automatic translation of animal sounds should be examined with caution. Artificial intelligence can help identify important patterns; understanding their meaning also requires evidence from the behavior and sensory perception of the person receiving the message.
Questions and Answers
Did the study examine bat sounds?
No. The researchers used toddler voices as an example of a communication system in which the context of some of the calls can be identified. The possible application to animal research is a methodological conclusion from this example.
Is artificial intelligence useless in studying animal communication?
It is useful for identifying and sorting patterns in recordings. The article warns that sorting sounds alone is not enough to determine what they are saying to the listener.
How can we examine the meaning of reading?
Recording and computer analysis can be combined with observation of the situation in which the call was heard, a controlled playback experiment, and measurement of the animals' response. Sometimes, research into their sensory and nervous systems is also necessary.
The scientific article
Taub and others, The challenge of decoding animal communication using AI, Current Biology (2026)
