X

Can autism really be detected by voice alone?

LenaBaby, a home device that records 16 hours of audio from a toddler's shirt pocket, appears to predict autism in children as young as 24 months with 91 percent accuracy.

Elizabeth Armstrong Moore
Elizabeth Armstrong Moore is based in Portland, Oregon, and has written for Wired, The Christian Science Monitor, and public radio. Her semi-obscure hobbies include climbing, billiards, board games that take up a lot of space, and piano.
Elizabeth Armstrong Moore
2 min read
The Lena System, a home device that records 16 hours of audio from a toddler's shirt pocket, appears to predict autism in children as young as 24 months with 91 percent accuracy. Lena Foundation

The Lena Foundation, whose new autism-screening tool hit the market in September, claims that parents who use the Lena System are now able to determine with 91 percent accuracy whether their child is developing normally, has autism, or has unassociated language delays.

The home kit, which includes a digital audio recorder, an outfit to hold the recorder, and a questionnaire about the child's development thus far, costs $699. (The one-time language and autism screen, on the other hand, is $200.) The foundation, which develops technology for the screening of several types of language delays and disorders, says the kit works for children as young as 24 months.

"Roughly speaking, autistic children vocalize differently from other children," Dongxin Xu, manager of software and language engineering at the Lena Foundation, tells MIT's Technology Review.

Analyzing a child's vocal patterns to screen for autism isn't new. The three factors that seem to set the Lena System apart from traditional screening methods are portability (the recording device is small); amount of data (16-hour recordings); and the software Lena uses to analyze the recordings parents mail in dutifully each month.

According to Jeffrey Richards, a statistician and database technician for the Lena Foundation, the software first categorizes the 16-hour audio stream into sound types, such as child, parent, or television. The child clips are then further dissected, and analyzed for the phonological composition of each sound, as well as how it is clustered and paired. The resulting data is then compared with the data compiled on children who are considered normal, autistic, or delayed.

The 91 percent accuracy is high, and while Lena researchers continue to fine-tune their software to push that rate even higher, I remain somewhat skeptical that voice alone can determine whether a child is autistic. It is often suggested that Einstein didn't speak until he was at least 3, if not 4 or 5; I have to wonder how a 16-hour recording of Einstein at 24 months would be interpreted by Lena software.