Initially, the UCLA team was developing a program to help hard of hearing parents, which could distinguish the child’s crying against the general background and give a signal. The application was called Chatterbaby and during its testing it became obvious that all infant sounds can be attributed to three categories: hunger, pain and boredom. An experienced mother can understand by intonation what the kid wants, but how about creating a software algorithm for this?
The task turned out to be simpler and more difficult than it was thought. As the infant acts reflexively, the pain senses make him scream continuously, while in the usual bustle between sounds there are distinct periods of pause. The food requirements have their own acoustic color and after studying 2000 samples the program has learned to recognize three types of screams with an accuracy of 90%. The question is only in the initial calibration of the system.
The genotype, exact age, climate, environment and other factors strongly affect the voice of the baby, so at UCLA they decided to compile a database with samples of screams of as many children as possible. Chatterbaby users are encouraged to record and send the votes of their offspring, but the information will not be personalized. It’s just technical data, tests for system debugging and optimization algorithms.
But such a bank of infant voices may have a wider application – for example, early recognition of autism. In elderly people with this problem, the voice changes significantly – maybe after analyzing thousands of records, it will be possible to deduce such patterns for babies as well? For this, parents are asked to go to the researchers to meet them and send short, for 5 seconds, records of their children’s voices, plus regularly undergo comprehensive surveys so that they can compare the data and conduct observations.