The best way to think about littleWords speech app is through the child’s comfort, the family’s real routine, and communication support that does not become pressure to perform. Home practice works best when it stays respectful and doable.
This is a builder story. If you want parenting advice, scroll past. If you want the story of a dad solving his kid’s problem with code, stick around.
The moment that changed the project
It was a Thursday in March 2024, maybe 6:45 p.m. My daughter Noa, four years old, was sitting cross-legged on the living room rug in our apartment in Austin with the prototype running on a propped-up iPad. My wife was making dinner. I was pretending to read on the couch but actually watching Noa’s face.
She’d been quiet all day. Her preschool teacher had flagged another rough afternoon: minimal verbal output, some frustration at circle time. Noa is autistic with a moderate expressive language delay, and rough days were not unusual.
Then she started talking to the screen.
“Buh-dee,” she said. Buddy, the character voice. “Buh-dee, I got a… I got a raff.”
Eight seconds of silence. The app waited. I held my breath.
“A raff. He’s yeh-yoh.”
Buddy responded: “A yellow giraffe! That sounds so cool. What’s your giraffe’s name?”
Noa grinned. Actually grinned. “George,” she said, clear as a bell.
Twelve minutes. She talked to that app for twelve straight minutes, which was four times longer than she’d ever engaged with any structured speech activity. She told Buddy about George the giraffe. She told Buddy about a Bluey episode. She used a word, “cracker,” that she’d been drilling with her SLP for six weeks, and the app just responded to it like a normal part of conversation instead of marking it correct.
I sat on that couch and my throat went tight. Not because the technology was brilliant. Because I had spent two years searching for the thing that would meet my daughter where she was, and it turned out to be a thing I could build myself.
The backstory, briefly
Noa got her autism diagnosis at two and a half. We did everything the system offers: Early Intervention, a private SLP twice a week at $175 a session, the waitlist for the developmental pediatrician (four months), the special-needs preschool. None of it was wrong. None of it was enough.
Somewhere around her third birthday I noticed something specific. Noa would talk to her tablet, sometimes a lot, when she would barely talk to me. Not because she didn’t love me. Because the tablet had no expectations. It didn’t try to read her face. It didn’t gently model the corrected pronunciation. It was, in some way I couldn’t quite name yet, safer.
I’m a software engineer. Fifteen years building products for adults. I had never built anything for children, and I had certainly never built anything for neurodivergent children. But I started prototyping anyway, because I had a problem in front of me that nobody was solving the way I needed it solved.
What most speech apps get wrong
Most speech apps for kids fall into two buckets: flashcard drills or gamified vocabulary trainers. Both fail autistic kids in similar ways. Drills create performance pressure. Gamification creates dopamine spikes that override the actual language goal. The kid chases stars and coins, not words.
What I wanted was simpler. A patient, friendly conversation partner that would wait for Noa as long as she needed, accept her approximations, follow her interests, and never grade her.
In product terms: a low-stakes, infinitely patient, prosodic conversational interface tuned for a kid with sensory and processing differences.
I tried building this in early 2024 with what was available. The first version was terrible. The voice was flat. The latency was too high. The model would interrupt before Noa finished thinking. It would correct her grammar (the opposite of what I wanted). It would ask follow-up questions in a way that felt like a job interview for a four-year-old.
I rebuilt it three times. The fourth time, it worked.
Five engineering decisions that actually mattered
If you’re a builder thinking about something similar, here’s what ended up making the difference. Not the model architecture or the cloud provider. These five design choices.
Long pause tolerance. Conversational AI defaults to detecting end-of-speech aggressively because adults expect snappy back-and-forth. I tuned ours to wait up to eight seconds of silence before responding. Eight seconds is an eternity in normal conversation. It is exactly right for an autistic four-year-old who is still assembling a sentence in her head. This single change was the biggest driver of Noa’s engagement. Nothing else came close.
Approximation acceptance. Off-the-shelf speech recognition tries to map kid speech to the closest adult English word and confidently gets it wrong. I added a custom layer that accepts approximations as the word the kid intended, without correcting back. If Noa says “ca-keh” for cracker, the app treats that as cracker and responds to cracker. The correction comes implicitly, from hearing the modeled word in Buddy’s response. Not from a red X on the screen.
Prosody over flatness. Most text-to-speech in kids’ apps is robotic in a way adults filter out but kids hear as wrong. I spent weeks tuning Buddy’s voice for rhythm, pitch variation, and emotional warmth. Kids learn the music of language before they learn the words. If the music is off, the words don’t land. Think of it like this: you can hand someone sheet music for a song they’ve never heard, or you can hum it for them. Buddy hums.
No scoring, no levels, no streaks. I deleted every gamification mechanic. No XP. No progress bar. No “you did great!” Just a conversation. Here’s the thing: autistic kids in particular are sensitive to extrinsic-reward structures. The moment a kid associates a tool with a quiz, the tool stops working. The willingness to talk evaporates.
Kid data never stored or sold. This was a baseline requirement, not a feature. The product is COPPA-compliant from the architecture up. Kid voice data is processed and discarded. No cross-user analytics on a per-kid basis. I have a daughter. I would not put her voice in someone else’s data pipeline. I assume you wouldn’t either.
None of these decisions were obvious from the AI tutorials floating around at the time. All five came from the same place: sitting next to Noa, watching her use the early versions, and asking our SLP, “What’s going wrong here?”
Four months of real use
Over the four months after that Thursday in March, Noa talked to Buddy for about ten minutes most evenings. Not every evening. Some days she wasn’t interested. Some days she just wanted to watch Bluey and eat Goldfish, and that was fine. The app was there when she wanted it.
Her SLP started commenting on increased conversational turn-taking during sessions. Her preschool teacher flagged more spontaneous speech at school. I won’t claim the app caused those changes in isolation. A lot was happening at once: therapy continued, she was maturing, we adjusted some sensory supports at home. But the LittleWords speech app gave her a real ten-minute window each day that she looked forward to, and that window did real work.
The boring truth about speech development is that volume matters. Kids need hundreds of low-pressure conversational turns to build fluency, and most families can’t manufacture that many turns organically during the chaos of a weeknight. Buddy could.
Why it became a product
About six months in, I started showing the prototype to other parents. Three things happened.
First, parents didn’t say “let me know when it launches.” They said “can I have a build right now.” There’s a difference. The first is polite interest. The second is someone who has been looking for this thing and just found it.
Second, SLPs asked to recommend it to families. Some of them had been cobbling together homemade tools for years and recognized what I’d built as something they could actually point parents toward.
Third, the results weren’t unique to Noa. The kids of the other parents I gave early builds to showed similar patterns: longer engagement, more spontaneous utterances, less resistance to practicing speech. The model generalized.
Once all three of those things were true, I had a choice. Keep it as a hobby project for our family, or ship it as a product. I chose product, because I’d spent two years feeling isolated in the autistic-kid-speech-delay world and I knew there were thousands of dads (and moms, and grandparents) who needed exactly this.
Pricing: $49 lifetime during the Founding Family waitlist phase. I made it lifetime on purpose. I do not want this to be another monthly subscription that families have to justify against the financial-planning anxiety of raising an autistic kid. Families already pay enough for therapy. They should not also be paying twenty bucks a month for an app they already needed.
(Honestly, I think the SaaS-ification of disability tools is one of the more cynical trends in tech right now, and I didn’t want to be part of it.)
What the app is not
I want to be precise about what I built and what I didn’t.
The app is not an AAC replacement. If your kid uses or needs augmentative communication, this is not that. Different tool, different purpose.
The app is not a substitute for SLP intervention. If your kid is in speech therapy, stay in therapy. Buddy is a practice partner, not a clinician.
The app is not a diagnostic tool. If you suspect your child is autistic or has a speech delay, get a professional evaluation. An app cannot do that, and anyone who tells you otherwise is selling something dangerous.
What the app is: a low-pressure speech practice companion for autistic and speech-delayed kids, roughly ages three to ten. One tool in a stack of tools. The tool I needed and could not find. The tool I’m offering to other parents who are standing where I stood two years ago.
The real takeaway for builders
If you’re reading this on an indie hacker site, here’s the part that matters to you.
The reason you can build a better tool for your own kid than any company can is that you live with the use case. You know what works at 6:30 p.m. on a Tuesday in February when everyone is tired and the sensory bucket is full and dinner is half-made. A product team shipping a polished demo from an office in San Francisco cannot know what that hour feels like. You can.
Build the thing your kid needs. If it works, give it to other parents. The market for parent-built tools for neurodivergent kids is enormous and wildly underserved. We are the people best positioned to fill it, because we’re the ones who go to bed thinking about it.
That’s the whole story. A dad. A daughter. A speech delay. A laptop. Eighteen months. One app. A growing number of families using it now.
The next one of these stories should be yours. I mean that. We need more of them.
