Building this as a solo dad means I have to keep the scope small, but the core experience is specific: a grown-up enters a child's name, age and wish, then an animated Santa video call made with AI says the child's name and talks about the wish. The aim is a personal Christmas moment, rather than a generic clip with a name pasted on.
The difficult part is making that feel natural and dependable. Names are not just text: spelling, pronunciation and accents can all make generated speech sound odd. I'm exploring how to make pronunciation easier to guide and how to handle a result that still sounds wrong. Because this is for children, safety has to shape the experience. I'm aiming to keep the interaction adult-led, with a bounded Christmas conversation rather than an open-ended chatbot; I'm also thinking carefully about the information collected and what a family should be told about AI-generated media.
It also needs to work beyond my laptop. On phones and tablets, layout, touch targets, audio playback, camera framing and variable network conditions all change the experience. Testing on real devices is as important as getting the desktop version working.
I'm building a few small browser Christmas games alongside the call. Families can try the games free for three days. If they choose a personalised call, the price is a simple one-off from £4.99, with nothing that renews. No subscription.
I'm building this independently, so feedback from people who've worked on speech interfaces, child-safe products or responsive web apps would be especially helpful. What would you test first, and where would you set the boundaries for generated dialogue?
You can see it at https://santaplaytime.com, and feedback is very welcome.
Top comments (1)
The bounded conversation decision is the right one, and I'd push it further than you might expect. For a child-facing product, the "open-ended chatbot" model is where every horror story starts — not because the model is malicious, but because it's inconsistent. A Christmas conversation has maybe twenty plausible branches, and constraining the model to those branches makes the experience better, not just safer. Kids notice when Santa says something weird.
On pronunciation: the hard cases are usually not the names you can predict. Anglo names are mostly fine, but names with digraphs (Siobhan, Niamh), names with non-obvious stress (Polish, Hungarian), and names where the spelling is ambiguous across languages will break a phonetic hint system. The practical fix I'd suggest is a "say it how you say it" step for the parent — a short field where they can type a phonetic respelling, plus a test-play button so they hear it before committing. Costs almost nothing to build, and it turns a failure mode into a feature.
The one thing I'd test first: audio playback on iOS Safari, specifically whether the generated audio survives the transition from a user gesture to an asynchronous fetch. Autoplay restrictions on Safari are stricter than anywhere else, and if the child taps the screen and nothing happens for three seconds, the magic is gone. If you can make that feel instant, everything else is polish.
Where I'd set the dialogue boundaries: nothing that asks the child for information beyond what's already on the form. No "what's your address so I can send a present," no "what's your school like." The adult enters the data; the child only listens and maybe picks from two or three options. That keeps the safety story simple — the child never inputs anything, so there's nothing to leak.
Good luck with the launch. The £4.99 one-off is a smart price point for a one-time emotional purchase.