Babies’ cries alone don’t convey what they want
New research finds that neither trained participants or AI can distinguish what babies need from the sound of their cries.
12 April 2024
By Emma Young
For every new parent who's been instructed by a relative that their baby's cry clearly means that it's hungry — or tired, or maybe needs its nappy changing — this finding is for you: according to new research, babies' cries contain no clue to their cause. In fact, Marguerite Lockhart-Baron at the University of Saint-Etienne, France, and colleagues report in Communications Psychology that neither artificial intelligence nor specially trained people could identify the cause of a baby's crying from the sound.
One reason this new finding is important is that it's not just family members who will often claim to be able 'read' a baby's crying. "Some non-academic sources even suggest that babies' cries are a 'language' made up of phonemes whose meaning can be learned," the team writes. Phone apps that promise to 'translate' cries are also becoming increasingly popular — "despite a lack of fundamental scientific evidence to support their veracity."
A problem with earlier research, though, is that it's generally been small-scale, making it hard to draw clear conclusions. In a bid to address this, Lockhart-Baron and colleagues created data set of almost 40,000 crying sequences extracted from 48-hour-hour long recordings of 24 babies (10 girls and 14 boys) in their homes. These recordings were made when the babies were aged 15 days, 1.5 months, 2.5 months, and 3.5 months.
Whenever a baby cried during a recording phase, a parent noted when the crying started, what they thought the reason might be, and also which action was effective at soothing it — which the team used to indicate the cause. In total, the researchers acquired 676 snippets of crying for which they knew the cause was hunger, discomfort, or isolation.
Next, they took two thirds of the crying sequences, and gave a machine learning algorithm scores for each of 10 acoustic variables (such as pitch and 'roughness') plus the cause. This was the AI training phase. In the testing phase, they gave the algorithm the acoustic data for the remaining one third of samples, and asked it to predict the cause. The results showed that it did no better than chance. In other words, there were no links between any particular patterns of these 10 acoustic variables and the reason for the crying.
The data also showed, however, that about three quarters of the time, the parents' predictions about the reason for the crying turned out to be correct. To explore whether people are able to pick up on something in the sound that wasn't captured with the ten acoustic variables, the team recruited 146 adult listeners who listened to a set of cry sequences from the babies when they were all 1.5 months old, and who were asked to pick a cause each time.
Again, they did no better than chance. (The team also found little to no difference in the performance of women versus men, or parents versus non-parents.) This suggests that the parents of the babies whose cries were sampled were often right about the cause thanks to their background knowledge — about how long it had been since the last feed or nappy change, for example — rather than because of any information contained in the cry itself.
In a final study, the team tried training adults, instead of AI. These participants had to guess the cause of a series of cry samples from a single infant, but were given feedback after each answer. After this training, they were played further sequences from the same baby. Once again, when asked to identify the cause, they did no better than chance.
As ever, this study does have a few limitations. One is that, with age, most babies cry less, so the team had relatively few recordings of crying from the babies at 3.5 months old. For this reason, they don't rule out the possibility that the crying of babies of this age — or older — may contain clues to the cause that are absent from the cries of younger babies.
Given that, in theory, it could be useful for a baby's cry to signal the reason, it might seem surprising that the team did not find this. But they did also find (in agreement with results from some earlier studies) that each individual baby had a 'cry signature', which persisted over time. (The most significant factor for this was median pitch. More than any other variable, this was useful for identifying a single baby.)
A baby's cry can contain only so much information. The researchers suggest that from an evolutionary perspective, perhaps it was more useful for a baby's cry to clearly signal its identity rather than a specific category of distress — which might have muddied the cry signature.
One key takeaway from the new study, though, is that if a family member knows when a baby was last fed or had its nappy changed, they might have some insight into why it's crying. But, if they claim to know the cause from the sound alone, this study at least enables you to confidently instruct them that they're wrong.
