A free Hebrew/English Android game for toddlers (ages 2–4), with פוקי, a talking parrot. No cloud, no ads, no tracking, and nothing about your child ever leaves the device. Open source, under the MIT license.
Website: pukiparrot.com (Hebrew and English, with the latest download)
I made Puki because I got tired of toddler games that stop in the middle to ask for money. My daughter would get into a game, love it, and then a screen would pop up: pay, watch an ad, unlock the next level. A two-year-old can't understand any of that. All she learns is that the fun suddenly ends.
So Puki is free, forever: no price, no in-app purchases, no subscriptions, no "full version", no ads. It will never ask for money, and there's no payment code in it at all. The only thing that ends the fun is the time limit the parents set, and even that ends with a calm goodbye. It's open source (MIT), so nobody can ever lock it behind a payment.
The game: an island world with toddler-sized games, each in its own illustrated scene (a farm, a garden, a playground under a rainbow, a theater stage, a cozy house):
- Animal Garden: feed each animal the right food. The animal munches and does a happy hop.
- Colors, Shapes, Counting, Letters: pop the right balloon, drag chunky shapes into place, count flowers into a garden bed, find the letter block, all with spoken Hebrew prompts. Nothing stands in rows: balloons float high and low, shapes lie around the target, and every round is arranged a little differently.
- Music: a drum, a xylophone, a tambourine and maracas on a stage, each with its own sound (the xylophone walks up and down a scale), then repeat a little tune.
- My Day: a cutaway house with a scene for each routine: soap the muddy spots off a child in the bath, set the plate, cup, fork and spoon on a table, dress the child from an open wardrobe, put the toys away around a room, brush Teddy's teeth (toothpaste, scrub, swish), bake a cake in the kitchen, wash a muddy car in the garage, and bedtime.
- Drawing on paper taped to an easel, with big paint pots, an eraser, stamps, color-in pages and tracing, and a camera that saves the picture. A sticker album fills up as stars are earned.
- Peekaboo barn: an animal hides on the farm (in the barn, behind the bush or the haystack, in a tree stump) and its sound gives it away. Tapping the right spot brings it out; any other spot sends out a butterfly and the animal peeks and calls again.
- Echo Puki: the child says anything and Puki says it back in a silly parrot voice (her own recording, played higher and faster). No brain needed, and nothing is kept.
- Say it with Puki: a picture appears, Puki says the word, and the child says it back. With the device brain on, the word itself moves the game on: after another word Puki says "listen well" and the word again, after a near miss "almost", up to three tries, and Puki always closes by saying the word the right way. Without it, any try counts.
Toddler-first design: no reading needed, big tap targets, no failure states (wrong answers get a gentle shake and "let's try again"), and every line spoken in a warm voice. When the parent's time limit is up there's no countdown: evening falls over the game, Puki yawns and waves goodbye, and the music fades. Everything answers a touch: pieces squish when pressed, bob while they wait, and burst into stars when they're right. Hebrew instructions follow the child's gender (תלחצי / תלחץ), set in parent settings. All voice lines are generated offline with BlueTTS.
Puki works in one of two ways:
- Recorded lines (the default). Everything in the game is spoken from recorded lines, and the parrot answers taps with recorded giggles and squawks. Nothing to download.
- The device brain. Puki writes his own lines about what the child just did, speaks them, and listens when she talks to him. Holding the parrot lets her talk to it; after a question ("which color?", "how many?", "who's hiding?") he listens for a few seconds, so she can say the answer instead of tapping it. It speaks Hebrew only for now; in English the game uses its recorded lines. Drawings and bedtime stories use recorded lines for now too.
The device brain is a one-time download of about 0.5 GB (538 MB) from Hugging Face. A parent starts it from the Puki on this device card in parent settings. After that it works offline. On a Galaxy Tab A7 Lite a reply takes about 1.5 s, from what happened to Puki speaking.
It's four small models, all ONNX int8:
| Model | What it does | Where it comes from |
|---|---|---|
| Brains | Writes Puki's lines, and rates what the child said (the safety gate) | A 46M-parameter model trained from scratch for Puki: pretrained on 1.6B tokens (about 74% Hebrew), then fine-tuned on 46k of Puki's lines. See train/. |
| Beak | Reads plain Hebrew into phonemes, so the voice says it right | 2.3M parameters, trained by us: pretrained on 600k sentences labelled by RenikudPlus, fine-tuned on reviewed lines |
| Syrinx | Puki's voice | BlueTTS with the Noa voice, its vector estimator and vocoder rewritten and quantized to int8 for weak tablets. Not trained by us. |
| Ears | Hears the child, in Hebrew and English | kol-stream by MaorB, exported for phones. Not trained by us. |
Every file is pinned by size and SHA-256 in the app (DeviceModels.kt), so a changed file is
refused. A parent can delete the models again from the same card.
- What the child says is checked first. Word lists catch hurt, danger, wanting a parent, big questions and secrets, and then Brains' own gate rates what she said. Either one gives a fixed, gentle reply that points to mommy or daddy; the model never writes those. If the gate isn't sure, Puki answers with a neutral line.
- Every line Puki writes is checked after. A filter lets through only short, clean lines, with no promises ("mommy is on her way") and no bonding ("best friend", "I love you"). A line that fails is dropped, and the game plays a recorded line instead.
- Parents see what Puki said. In parent settings, "What Puki said" lists his recent lines. A parent can flag one that wasn't right; the report goes by the parent's own email.
- Android 8 (API 26) or newer. Built for tablets in landscape; phones work too.
- For the device brain: a 64-bit ARM device with about 3 GB of memory, and about 0.5 GB of free space. The card only appears on devices that can run it.
- Nothing about the child leaves the device. No accounts, ads, analytics or cloud calls.
- The voice is understood on the device, and no recording is kept. Echo Puki's recordings go straight to the speaker a moment later and are never stored or sent.
- The child's words are held in memory only for the parent's "What Puki said" list, and are forgotten when the game closes.
- The brain keeps no memory beyond its last few lines.
- The child's name lives only on the device. The files that would hold it in this repo are git-ignored.
- The network is used for two things only: the model download a parent asks for, and, in the GitHub build, about once a day a check whether there's a newer version (a plain list request, with nothing about the child in it).
- Signed releases: every APK on the releases page (from 0.4.1) is signed with Puki's release
key, certificate SHA-256
a799cab5bcf83c20c754490f8d0eb271cf844e2b7e69518760789f2a96eb3c3d. Android only installs an update signed with the same key as the installed app, so a changed Puki can't come in as an update. To check a download:apksigner verify --print-certs Puki-<version>.apk.
- APK: download the latest
Puki-<version>.apkfrom the releases (or the website) and open it on the tablet. - Obtainium: add
https://github.com/TomSadeh/pukiin Obtainium to get each new release. - Google Play: coming soon.
First launch: a short welcome screen asks for the language, the child's name, and girl/boy.
Parent settings: long-press the gear on the hub, then tap the number written out in words (a toddler can hold a button, but can't read "seventeen"). There you can change all of the above, turn on the device brain, see what Puki said, and set separate volumes for music, voices (and Puki) and sound effects, which games show on the map, and a time limit. The limit holds even if the app is closed; a new allowance starts after an hour away, or when you end the goodbye.
Updates: the GitHub build updates itself from this repo's releases. When there's a newer version, a red dot shows on the gear; in the parent settings, tap Update. The first time, Android asks you to let Puki install apps (a switch on the next screen); after that, each update just asks you to confirm. Progress, stars and settings are kept.
Personal voice lines: the game can say the child's name ("כל הכבוד, ___!"). When the device brain is on, Puki records these lines onto the device (in Hebrew, for now), so they play at once, even with the brain off later. The name never leaves the device.
Until 0.8, Puki's brain could run on a PC at home. In 0.9.0 the PC brain is retired: the app no longer talks to a PC, and the device brain does its job. If your family used it:
- Update the app to 0.9.0 or newer.
- Turn on the device brain: parent settings, Puki on this device, then download.
- Remove the PC program. It installed a login shortcut, a firewall rule, and about 10 GB
of models. To remove all of it:
- Puki.exe (Windows): right-click Puki's tray icon (behind the ^ arrow if you don't see it), choose Uninstall Puki..., and answer yes to each part. Then delete Puki.exe.
- Installed from this repo: run
python brain\install.py --uninstallon Windows, orpython3 brain/install.py --uninstallon Linux (add--dir <folder>if you installed it somewhere other than~/puki-brain). It stops Puki and asks before removing each part: the login shortcut (or thepukisystemd user service on Linux), the firewall rule, the models, and your settings and pairing code. - By hand: quit Puki; delete the "Puki brain" shortcut from
shell:startup; in an administrator PowerShell, runRemove-NetFirewallRule -DisplayName "Puki companion"; then delete the folder it ran from (~/puki-brain, or%LOCALAPPDATA%\Pukifor Puki.exe).
The details are at the top of brain/README.md.
app/ Android app (Kotlin, Jetpack Compose)
app/.../puki/brain/ the device brain: LocalBrain (routes), Brains, Bpe, Persona, Safety,
KolEars, SpokenAnswer, DeviceModels (pinned downloads), voice/ (Beak,
BlueVoice)
train/ how Puki's models are made (see train/README.md)
brain/ the retired PC brain (Python); safety.py, persona.py and
pronunciation.tsv are still the source for the app and train/
scripts/generate_voices.py all voice lines, generated with BlueTTS
scripts/check_voices.py checks every voice line the app plays has audio
scripts/sync_safety_kt.py copies safety.py's word lists into Safety.kt (--check compares)
scripts/safety_eval.py the safety stress test
scripts/generate_instruments.py the music game's instrument sounds
scripts/release.py builds the APK and publishes it as a GitHub release
scripts/draw_companion.py draws the parrot
CLAUDE.md conventions for working on the code (also read by Claude Code)
./gradlew installGithubDebug
- BlueTTS and nikud: BlueTTS's Hebrew pronunciation model garbles text that already has
nikud, so all text is fed to it plain. Words plain spelling can't disambiguate (letter
names, animal sounds) are given as phonemes in
generate_voices.py. On the device, Beak reads plain Hebrew into phonemes. - The device brain (
app/src/main/java/io/github/tomsadeh/puki/brain/):LocalBrainanswers the game's requests (Brain.ktis the interface) withBrains(Puki's lines and the gate, tokenized byBpe),PersonaandSafety(the filter and nets, ported frombrain/persona.pyandbrain/safety.py),voice/(Beakreads,BlueVoicespeaks: a Kotlin port of BlueTTS's pipeline) andKolEars(kol-stream through ONNX Runtime). The benches inandroidTesttime each model on a device:gradlew connectedGithubDebugAndroidTest -Ppuki.benchinstalls them as a separate app. - Music: "Puki main theme" plays in the hub and games, and "Puki nights" plays during the
bedtime routine. Dance and Freeze plays
music_dance.mp3when it's there (otherwise the main theme). To use your own, replaceapp/src/main/res/raw/music_hub.mp3,music_night.mp3andmusic_dance.mp3. - Personal voice lines in the build: to bake the child's name lines into the app instead
(they're git-ignored, so the name never ends up in a repo):
python scripts/generate_voices.py --bluetts-dir <BlueTTS folder> --child-name "<name>" --child-gender girl - Releasing: bump
pukiVersioninapp/build.gradle.kts, commit, then runpython scripts/release.py --publishfrom this checkout. It refuses if personal voice lines (voice_*_name*) are inres/raw, so a release never carries a child's name, and if the APK is signed with the debug key (releases use the key in~/keys, seeapp/build.gradle.kts). - Tests:
./gradlew testGithubDebugUnitTest,python -m pytest brain/tests,python scripts/sync_safety_kt.py --checkandpython scripts/check_voices.py(CI runs them all). On a device, the debug build's self-test:adb shell am broadcast -n io.github.tomsadeh.puki/.SelfTestReceiver(adb logcat -s PukiSelfTest), with-e timelimit 1to see the goodbye after a minute (0turns it off again), or-e echotest 1for the parrot voice. - Changing a model: put the new files' sizes and SHA-256 in
DeviceModels.kt, updatetrain/pt_manifest.jsonif the pretraining changed, and re-run the truth and safety checks (train/sft_small.py's held-out scoring: the filter, the gate and the truth checks) before a release. - Adding a voice line: add it to
voice_linesinscripts/generate_voices.py, generate it, then runscripts/check_voices.py.
- Music ("Puki main theme", "Puki nights" and the Dance and Freeze track) by the project's author, Tom Sadeh: piano, produced with Suno; CC BY 4.0
- AndroidX and Jetpack Compose by Google (Apache-2.0)
- ONNX Runtime by Microsoft (MIT) runs all four models
- BlueTTS by Max Melichov (MIT): Puki's voice, with the Noa voice, and the recorded lines
- RenikudPlus by Max Melichov, Yakov Kolani and Morris Alper (Apache-2.0): Beak learned to read Hebrew from it
- espeak-ng (GPL-3.0): English Beak learned to read from its pronunciations (only its output; espeak-ng itself isn't part of Puki)
- kol-stream by MaorB (Apache-2.0), on Google's EmbeddingGemma 2, trained with ivrit.ai's data (CC BY 4.0) and LibriSpeech: Puki's ears
- The Hebrew Brains' pretraining text: FineWeb-2 (heb_Hebr) by Hugging Face (ODC-By, and the Common Crawl terms of use)
- The English Brains' pretraining text:
- FineWeb-Edu by Hugging Face (ODC-By, and the Common Crawl terms of use), which also gave English Beak its words
- Cosmopedia by Hugging Face (Apache-2.0)
- Project Gutenberg books and YouTube Commons videos from the Common Pile (public domain, and CC BY per video)
- Brains' fine-tuning lines were written by DeepSeek models, then filtered by Puki's own checks; they're public as dtsj89/puki-lines (CC BY 4.0)
The models are downloaded separately and are not part of this repository.
Google Play, a few more pictures for Puki's games (a doll, a blanket, the sea), and whatever families ask for. See ROADMAP.md.
Pull requests are welcome. The rules come from the audience: no failure states, nothing the child has to read, big targets, both languages, and nothing about the child leaves the device. See CONTRIBUTING.md for what gets accepted and what a PR needs to show. Porting the brain to another platform (iOS, say)? PORTING.md describes the models and how to check a port against the trainer's fixtures. Questions, or something Puki said that wasn't right: hello@pukiparrot.com.
The code is MIT. The music is CC BY 4.0. Built with lots of help from Claude Code.