We studied what worked, and what failed, before we built it out.

What the studies say

The research base is thin, but it points one way: a grown-up in the loop, music, and no waiting on the machine.

Robots and AI engage. People teach.

The only controlled social-robot trial we found with minimally verbal autistic children (37 children who speak few or no words, ages 4–12, with a therapist present) found 159% more children improving on language measures than usual care. Requests per session rose from 6.8 to 16.7, with no adverse events. Whether the gains carried over to conversations with people, not just the robot, was never tested. (J. Clin. Med., 2025)

An AI chat study with 12 verbal autistic children got them to use 13% more words and talk 43% longer than with a human, but the human's responses scored 11% higher on quality. (ASD-Chat, arXiv) The one AI tool we found for minimally verbal autistic children helps a parent and child take turns talking with each other; a study of 11 parent–child pairs found gains in engagement and turn-taking. Its lessons: make turn-taking explicit, value the child's own words over "correct" speech, always offer a way to refuse, and use data to show what the child can do. (AACessTalk, arXiv)

What we built from it: a supervised play session that leaves the teaching to people, "something else" offered with every choice, "Do you want to play today?" asked first, and partner training.

Waiting is a skill. Lag is a problem.

Experts reviewing an AI plush toy flagged 6–10 second processing delays as anxiety-inducing (Echo-Teddy, arXiv). One published review of another AI toy reported 10–15 second delays and about 20% irrelevant answers when a child's speech was quiet or mumbled (Mommyhood101). But deliberate waiting can help: in the classic time-delay study, a gradually lengthened pause before an adult modeled a request taught all seven autistic children to ask on their own, in new places, with new people. (Claremont, 1985)

What we built from it: Buddy acknowledges an answer right away, before the AI replies, and the wait is adjustable: a quiet pause, a small "Yes… or no?" cue, the pause again, then a reminder. The first pause can grow from 3 to 5 to 8 seconds over the weeks.

Singing with a drum has the best evidence we found for speech sounds

Auditory-Motor Mapping Training (AMMT) sings two-syllable words on two notes while child and therapist tap drums. In a 15-child randomized trial with minimally verbal autistic children, it improved syllable accuracy by 12.1 points versus 2.8 for speech repetition therapy (an effect size of d = 0.82), with gains on untrained words too. Children with the most severe impairment benefited least. (AMMT trial, PubMed Central) Separately, a 2024 analysis of 18 music-therapy trials linked intensive schedules (about five sessions a week, over a few weeks) to larger effects, though its data have real limits (see below). (Frontiers in Psychology, 2024) A home or classroom app can make frequent sessions easy, alongside therapy.

What we built from it: Sing and drum, play inspired by AMMT, not the therapy itself. Buddy says a favorite two-part word on a high note then a low note, with a drum hit on each part; then it's the child's turn. Nothing is graded. Plus rhythm echo, a clap-back beat game.

Model it. Don't demand it.

The AAC field's consensus is that there are no prerequisites for communication. In AAC research, modeling words on a speech device, without expecting the child to use it, was linked to bigger gains in speech; as one summary puts it, AAC "does not delay speech… the reverse is true." (as summarized by Avaz and PrAACtical AAC)

What we built from it: Buddy is told to say "yes" and "no" often in its own speech and never to ask "Can you say yes?"

Dignity is part of the design

Age-respectful practice says a teen's interests shouldn't be judged by "age equivalents or cognitive level," and content shouldn't be recycled year after year. (PrAACtical AAC) Low-vision design guidance calls for dark backgrounds, one large target, strong borders and careful motion. (Boster et al., 2025) A critical review warns that generative AI can "reflect and amplify societal biases" about cognitive disability, and calls for designing it with disabled people, not just for them. (PubMed review)

What we built from it: age-respectful humor, celebration sounds a child picks for themselves, praise Buddy is told to keep natural, a dark, high-contrast screen with one big character, and a child's "no" that changes what happens.

What the evidence doesn't tell us

Ninja Buddy is built on what the research supports, not on proven outcomes.

• No study we found has tested finger snaps, or any non-speech sound, as a way to answer.
• No study we found has tested an AI that talks directly with children who speak very little, or included children with intellectual disabilities. (AACessTalk, described above, helps a parent and child talk with each other.)
• The robot and AI studies show more engagement, not proven communication gains from AI on its own.
• The AMMT evidence behind Sing and drum is about the accuracy of syllables and vowels, not speech a child starts on their own.
• The music-therapy analysis reports large effects, but they're weaker than the numbers suggest: its trials varied widely, showed signs that weaker results went unpublished, and nearly all were published in Chinese.
• Partner training, adjustable wait time and playing together follow the evidence, but none has been tested inside an app.
• Ninja Buddy hasn't launched yet, and it hasn't been studied. Its features are built and checked by automated tests, but most haven't been tried with children yet, so this site describes what each one is designed to do, not results.
• No recent study gives wait times for teens who speak very little, so Buddy's default waits are a starting guess you can adjust.
• Low-vision design guidance comes from work with young children; no study we found covers teens with both low vision and intellectual disability.

We hope to see more attempts, faster answers, and more answers that come before any cue. We don't promise speech. Every answer is logged so you and your child's therapist can see how answers change over time. Ninja Buddy works alongside your child's therapy and AAC, never instead of them.

Our research covered published products, studies and reports. It didn't include parent communities, so we'd love to hear from families, teachers and therapists directly before launch. Tell us what you think.