LM-Kit OneThe Private AI Application Server
Promoted
Launching today
StillTalk
A photo that talks back. Live on your device, not video.
27 followers
A photo that talks back. Live on your device, not video.
27 followers
StillTalk turns one generated photo into a character that talks. Your device draws every frame live in the browser; the server sends only speech and compact movement data. Five characters, six languages, or type your own sentence. No signup. Engine built with GPT-6 Astra.
Hi Product Hunt 👋 I'm Piotr, founder of StillTalk.
StillTalk turns one generated photo into a character that talks, and it is not video. Your device draws every frame live in the browser. The server sends only speech and compact movement data: no video stream, no cloud renderer billing by the minute.
Try it now, no signup: pick one of five characters, one of six languages, play a tongue twister, or type your own sentence and hear it spoken.
The Astra story: GPT-6 Astra designed and implemented the rendering engine, the articulation, the calibration and the tests. My job was to set the quality bar and keep saying "not yet": a blink sitting too high, a mouth shape that worked in Polish but not in French. We found this challenge on Saturday, September 12. The engine worked with one character. Over the next day Astra built four more, and two of them needed very few corrections from me, because every earlier fix had already become an automated check.
For scale: 1.8B+ tokens in the StillTalk thread alone, and 817M on our busiest single day. That includes reused context. It measures the work, not a bill.
Honest limits: live typing runs on a free speech pool with per-person limits. If it runs out, every prepared line still plays.
Next: a commercial API, a wider character library, and a beta for creating a character from your own photo.
This is the result of one very intense sprint. I'm happy with where it is, and I'm still calibrating every day, so it will only get better from here.
I'd love your feedback: which character feels most alive, and where does the lip-sync break in your language? I'm here all day.
The language support is interesting. Does the lip sync stay accurate when you switch between languages?
@sheikh_umair1 Great question, Sebastian. Yes, because nothing is shared blindly between languages: lip sync is calibrated on a speech corpus per language and per character, 980 samples across six languages so far. Switch Clara from English to Polish and the calibration switches with her. Today every character/language pair has its own tuning; we're unifying that now, and the goal is for the adjustments to adapt automatically to any new character, including one created from your own photo.
Why these six languages? StillTalk started as the avatar for our own language-learning app. When the challenge came up we thought: if we need this, others probably do too. More languages are planned.
The demo is live on stilltalk.ai, no signup: pick a character, switch languages, play a tongue twister and judge for yourself. If you spot a weak character/language pair, tell me which one. That's the most useful feedback I can get.
I could see this being useful for quick videos when you don’t want to record yourself. Pretty simple idea, but I like it.
@kyle_bennett6 Thanks, Kyle! One twist: it doesn't make videos at all. The character speaks live on the page. Type a sentence and it says it; nothing is rendered or exported. That's what makes it light enough to live inside an app or a website as a tutor, a guide or a support agent.
That said, export to video is an interesting idea for an extra feature. It's actually how we made our launch film: the engine played the conversation and we recorded it. Try typing your own line on stilltalk.ai and see what you'd want to export.
Honestly, the talking photo thing is kinda cool. I wonder how natural it looks with different photos.
@ayesha_mughal1 Thanks, Ayesha! Very natural, and across very different faces: five characters so far, from photoreal people to a cartoon creature, four of them built in about a day. Your own photo is next, and so is richer full-body motion: a wave hello is almost ready. Drop your email on stilltalk.ai and you'll see for yourself soon.
This could be fun for making little character videos. Curious what people are using it for so far.
@charlotte_reed1 Thanks, Charlotte! We launched today, so the honest answer is: our first use is our own language-learning app, where a character speaks with learners in six languages. Note it's not a video maker. The character talks live on the page, so it fits anywhere you'd want a tutor, a guide or a support agent that can say anything on the spot. The demo is open on stilltalk.ai: make Pistak say something and tell me what you'd use it for.