LINGUISTIC disorders of speech or of comprehension are awkward for anyone who suffers from them. For children, who are just beginning to make their way in society, they can be disastrous. Teasing, bullying, lack of friends and poor school performance may all follow from an inability to talk or listen normally. Early intervention and therapy, though, can make a big difference—if diagnosis comes quickly enough.
Often it does not. In America, 60% of such disorders go undiagnosed until a child goes to school. But Jen Gong and John Guttag of the Massachusetts Institute of Technology hope to change that. As they outlined at the Interspeech Conference in San Francisco in September, they have devised a method that, when refined, may yield an automated test which can spot the subtle clues, such as pauses during speech, that indicate a disorder to a professional ear but may not be obvious to parents.
Ms Gong and Dr Guttag, both computer scientists, wondered whether they could teach their machines to distinguish the speech of children with disorders from that of children without them. To this end, they first wrote an algorithm they hoped might do so, and then collaborated with two speech pathologists, Tiffany Hogan and Jordan Green of the MGH Institute of Health Professions, to test it. Together, the researchers recorded 231 children between the ages of four and 17 retelling a story in their own words while being shown visual prompts. Dr Hogan and Dr Green had previously identified 192 of these children as developing normally in matters linguistic, while 39 had, in the two experts’ opinion, a speech or language disorder.
Ms Gong and Dr Guttag then let their algorithm loose on the audio samples. After chewing on the files in question, it noted that many characteristics—including the number of pauses, variations in pause durations, and the ratio between pauses and distinct segments of speech—were useful for detecting the presence of language and speech disorders. Ms Gong reported to the conference that the system was able to detect 72% of the children diagnosed by Ms Hogan and Dr Green as having an impairment. It also had a fairly low false-positive rate, suggesting impairments in only 18% of children not so diagnosed by the two human experts.
Neither of those numbers is good enough for a clinical system, but they provide a starting-point for one. And if such a system were developed, it would easily be translatable into the sort of app parents might routinely use to test their children—and thus receive early warning if something is wrong.