Artificial intelligence for the measurement of vocal stereotypy

MHeald et al. (2020) trained an AI to listen for vocal stereotypy in kids with autism. They recorded eight children during regular sessions. A neural network learned to flag the repetitive sounds.

Human observers also scored the same tapes. The team then compared the AI counts to the people counts, session by session.

The AI matched the humans on at least eight out of every ten sessions for six of the eight kids. In other words, the computer agreed with trained staff 80% of the time or better for most participants.

The result shows a cheap laptop can do the tallying instead of a tired clinician.

Maharaj et al. (2020) did the same trick with a Kinect camera. Their system tracked motor stereotypy with 92% accuracy. Together, the two studies prove AI can watch or listen and still hit the 90% club.

Gilchrist et al. (2018) and Keintz et al. (2011) used small accelerometers on wrists or torsos. They caught hand-flapping and body-rocking at 80-90% accuracy. MM et al. extend this line by adding the audio channel, so now both movement and sound can be scored without human timers.

Shawler et al. (2020) looked at how to reduce vocal stereotypy once you measure it. Their RIRD treatment paper pairs nicely with MM’s measurement paper—you can let the AI count while you run the intervention.

You no longer need a second staff member glued to a stopwatch. Set a tablet on the desk, hit record, and the AI logs vocal stereotypy for you. Use the saved staff time to teach instead of tally. Start small: record one client for one session and check the AI count against your own. If the numbers line up, let the model carry the load next week.