Whether an AI toy is still being carried around three months after launch, or has been dropped in a storage bin, looks like a content problem. It is often something smaller: does it wake when the child calls, and does it stop talking when the child cuts in.
Start with power. A toy cannot run a full speech recognizer continuously, because the battery will not take it and because ambient audio should not be streaming to the cloud all day. The usual design is a tiny wake-word detector running on-device, scoring ambient sound every second and handing off to recognition, understanding and speech synthesis only when the score crosses a threshold. Vendors of offline voice chips publish numbers that make the case: tens of microamps in deep sleep, under 1mA while listening, and 1 to 5mA running command-word recognition at full load, which is roughly half a year on two AA cells. That is why mid-tier toys moved on-device this year: chips such as the ESP32-S3 accelerate ML with vector instructions by three to five times, and the RK2108 carries a 0.5 TOPS NPU, enough for wake words, pronunciation scoring and simple intent locally.
The Threshold Is the Hardest Setting
Loosen the wake threshold and a TV ad with a similar-sounding word sets it off. Tighten it and a softly spoken child two metres away gets ignored. The steadier commercial pattern is two stages: a local detector that fires a little eagerly, then a stricter second check, sometimes in the cloud, to confirm, so most false wakes are swallowed before the child notices. The cost is an extra delay, and that delay is exactly what the interaction feels like.
Published offline-wake measurements show the spread. In a quiet room at 30dB, a good implementation reaches around 99% wake rate with roughly 320ms of response. At 65dB of noise, some designs fall from 98% to 61%, while a model retrained on real children's speech holds about 89%. There is a common trap here: fine-tuning on an adult corpus leaves the toy mis-waking on a child shouting a similar phrase, because the training data never covered the high-frequency distortion of young voices. Fixing false wakes means retraining on children's own recordings, plus a microphone array for noise suppression.
Not Letting the Child Interrupt Is Worse Than Not Hearing
Interruption, or barge-in, is what lets a child stop the toy mid-answer and change the subject. Without it, the toy finishes the whole reply before accepting another line, so three quick questions come back as three queued answers and the toy reads as laggy. With it, the moment the child speaks, the toy stops and listens. It is harder than it sounds: the toy has to tell the child's voice apart from its own speaker output, which takes echo cancellation and a fast decision to cut the current reply. There is a direct test at the sample table. Ask one question, wait half a second, then ask a second while the first is still being answered. A design with real interruption stops and takes the new line; a design without it answers both, one after the other.
So when you pick an AI toy, do not stop at which model it runs. Ask three numbers: how many false wakes per hour in a normal home, whether interruption is on by default, and whether the wake decision happens on the device or in the cloud. The first two decide whether the child keeps talking to it. The third decides whether the child's recordings leave the room.
© SWPO. All rights reserved.