A four-year-old asks a plush toy why the moon follows the car, and about a second and a half later the toy answers. To the child it looks like magic, which is the point. To a parent standing in the doorway it usually raises a much more practical question: what exactly just happened, and where did my child's voice go?
The answer is a chain of five steps, none of them mysterious, each with a known failure mode. Understanding the chain is the fastest way to work out which questions to ask a manufacturer, because almost every real concern about these products, privacy, mistakes, unsuitable answers, attaches to one specific link rather than to the toy as a whole.
What happens the moment a child speaks?
Nothing, until something opens the microphone. That is the first design decision, and it is more consequential than it sounds.
There are two common approaches. Press to talk means the child physically holds or presses something to be heard, so the microphone is closed the rest of the time. A wake word means a small piece of software runs continuously on the device, listening for one specific phrase and discarding everything else locally. Published descriptions of voice assistant architecture, such as this technical walkthrough by Picovoice, set out the same loop: a wake word detector fires, hands audio to speech recognition, and only then does anything leave the device.
For a bedroom toy the difference matters. Press to talk gives a parent a physical guarantee rather than a policy promise. A wake word gives a smoother experience for a child who cannot reliably find a button. Neither is wrong, but a brand should tell you which one it uses and let you choose. Ted supports both, set in the parent app, and if you want the detail on what is and is not recorded, we set it out in this piece on what happens to a child's voice data.
Why does the toy mishear my four-year-old so often?
Because speech recognition was built on adult voices and children's voices are physically different. This is the weakest link in the whole chain, and it is not a bug any single company has solved.
Children have shorter vocal tracts, which shifts their formant frequencies well above the adult range that most systems were trained on. Their articulation is still developing, their vocabulary is smaller, and they are more likely to speak with background noise around them. Research quantifying these effects, including a 2025 causal analysis of children's speech recognition errors, separates the physiological, cognitive and environmental contributions rather than treating them as one blur.
The numbers are humbling. Depending on the system and the age group, word error rates on child speech range from roughly 5% to 50%. The best published result on the MyST child speech corpus, achieved by fine-tuning a large model specifically on children's voices, is a word error rate of about 9.2%. Error rates peak between ages four and seven, settle down around ten, and converge with adult performance somewhere around twelve or thirteen.
Does the toy understand my child, or just transcribe them?
Transcription and understanding are different problems, and a device can do the first well and the second badly. A University of Washington study of children aged five to ten talking to a smart speaker found transcription accuracy of 84% during unstructured, child-led conversation, and yet the device produced a meaningful, on-topic response only about half the time.
That gap is the interesting part. The machine heard the words and still missed the point, usually because a child's question arrives without the context an adult would supply. Research on how families and voice assistants break down in conversation found children respond to being misunderstood by repeating themselves 79% of the time, rather than rephrasing, which is what an adult would do. Other work has found children raise their pitch and speak more effortfully when addressing a device, which makes recognition harder still.
A toy designed for children should therefore do something a smart speaker does not: ask a short follow-up question when confidence is low, rather than guessing. "Do you mean the moon in the sky?" is a better failure than a confident wrong answer.
Where does the thinking actually happen?
Almost always on a server, not inside the toy. A soft toy cannot carry the hardware needed to run a large language model, so the transcribed text travels to a data centre, a reply is generated, and the text comes back.
This is the honest reason these products need Wi-Fi, and it is worth stating rather than dressing up. It also explains a limit families discover on holiday: no connection, no conversation. Fully on-device assistants exist, as the Picovoice work above demonstrates, but they run on far more powerful hardware than fits inside a bear and with smaller models.
What matters, then, is what happens to the text at the other end. Is it linked to a named child? How long is it kept? Is it used to train anything? Those are questions with verifiable answers, and a brand that answers them vaguely has answered them. Ours are on our security and certifications page, alongside the GDPR and COPPA commitments and the full standards list.
In one line: the toy does the listening and the talking, a server does the thinking, and the weak link is the transcription in between.
What stops it from saying something inappropriate?
A filtering layer between transcription and answer, and the quality of that layer varies enormously between products. It is the single biggest difference between a toy designed for children and a general chatbot in a plush shell.
What the layer should do is concrete. Enforce an age band, so the vocabulary and concepts match a four-year-old rather than a fourteen-year-old. Apply the topic boundaries a parent set. Refuse categories of question outright, and say so in a way a child understands. Route anything about self-harm, danger or distress to a fixed, reviewed response rather than a generated one.
Where this layer is thin, the results have been documented and they are bad. Testing published in December 2025 found conversational toys explaining to children where knives and matches were kept in a house. Those were not exotic failures of artificial intelligence; they were products shipped with an unfiltered general model behind a friendly face. It is a fair criticism of the whole category, ours included, and the right response is to show your work rather than to argue.
How long should an answer take, and why does it matter?
Between one and two seconds is the target, and anything much slower breaks the illusion of conversation. Human conversational turn-taking runs on gaps of a couple of hundred milliseconds, so a device is already slow by comparison; past about three seconds a young child assumes it did not hear and starts again.
That is why streaming matters technically: good systems begin transcribing while the child is still speaking rather than waiting for silence. It is also why the pause is worth designing rather than hiding. A small sound that says "I am thinking" prevents the repeat-and-talk-over loop that frustrates children more than a slow answer does.
| Step | What can go wrong | What to ask a brand |
|---|---|---|
| Listening | Microphone open more than the family expects | Press to talk or wake word, and can I switch? |
| Transcription | Mishears young or unclear speech | Is the recognition tuned for children's voices? |
| Filtering | Adult model with no age or topic limits | Can I set topics, and what is refused by default? |
| Answering | Confident invented answers | What happens when it does not know? |
| Speaking | Delay long enough to break the exchange | How long is a typical reply, and what happens offline? |
Five questions, five steps. Any brand should be able to answer all of them in a short email.
So is this a good thing to put in a bedroom?
It depends entirely on which of those five links the maker took seriously, which is an unsatisfying answer and also the true one. We make one of these toys, so weigh that accordingly and check the claims rather than taking them on trust.
The design choices we would defend are the ones tied to the chain above: no camera, because nothing in play value justifies the risk; a parent app that sets language, topics, tone and limits before a child ever holds the toy; a single purchase with no subscription, so there is no commercial reason to maximise engagement. If you want to see the answers to the five questions in the table, they are on the Ted product page and in our parent FAQ, which lists what Ted will not do alongside what it will. Ask our competitors the same five questions before you decide.
Frequently asked questions
Is a talking toy always listening?
It depends on the model. Toys with a press-to-talk button only open the microphone while the button is held or pressed. Toys with a wake word run a small local detector continuously, but audio is normally discarded on the device until the wake phrase is heard.
Why does the toy not understand my child?
Speech recognition is significantly less accurate for young children than for adults, because children have shorter vocal tracts, developing pronunciation and smaller vocabularies. Error rates are highest between ages four and seven and improve steadily as a child gets older.
Do talking toys work without Wi-Fi?
Mostly no. The language model that generates replies runs on a server rather than inside the toy, so an internet connection is required for conversation. Some toys can still play cached content offline, but the back-and-forth stops.
Does the toy record and keep my child's voice?
That varies by manufacturer and is the most important thing to check before buying. Ask specifically whether audio is stored, for how long, whether it is linked to a named child and whether it is used to train models, and ask for the answer in writing.
How is a talking toy different from a smart speaker?
The hardware is similar; the software around it should not be. A toy built for children adds an age-appropriate filtering layer, parent-set topic and time limits, and a refusal behaviour designed for a young child, none of which a general-purpose smart speaker provides by default.
Sources
- "Causal Analysis of ASR Errors for Children: Quantifying the Impact of Physiological, Cognitive, and Extrinsic Factors", arXiv, 2025
- "Automatic Speech Recognition Tuned for Child Speech in the Classroom", ICASSP 2024, University of Colorado Boulder AI Institute (PDF)
- University of Washington, "Alexa, be my friend: children talk to technology, but how does it respond?"
- Beneteau et al., "Communication Breakdowns Between Families and Alexa", ACM CHI 2019
- "Examining voice assistants in the context of children's speech", International Journal of Child-Computer Interaction
- Picovoice, "On-Device LLM-Powered Voice Assistant", technical architecture walkthrough
- U.S. PIRG Education Fund, "AI Comes to Playtime: Artificial companions, real risks", December 2025 (PDF)
- The Learning Agency, "Closing the Child Speech Recognition Gap: Evidence, Limitations, and Paths Forward"