AI companies publish "system cards" — technical documents explaining what their systems can and cannot do. They are written for researchers. Ours is written for you, because our users are children, and we think a parent's right to understand the AI their child talks to should not depend on reading a research paper.
What happens when your child speaks
Listening
Speech recognition (Google Cloud) converts your child's voice to text. It is configured for how bilingual children actually speak — Hindi and English mixed in one sentence, high-pitched voices, mid-sentence restarts. Child speech is the hardest class of audio there is, and it is the class we measure and tune for.
Understanding and responding
Large language models (from Google and Anthropic) work inside a tightly scripted frame: evaluate the Hindi your child produced — crediting the attempt, not punishing the English — plan Kiki's reply for the current lesson topic and your child's level, and note any error for later practice. The models never converse freely; they fill a bounded role we define.
Speaking
Kiki's voice (ElevenLabs) was not picked by taste. We ran an acoustic study across 15 candidate voices, measured their pitch, and chose a warm, Hindi-native voice in the register of an elder sister — because that is the register children's most beloved characters actually live in.
Why we built it this way
The fashionable way to build a voice AI today is a speech-to-speech model: audio in, audio out, nothing in between. We deliberately chose the older, more inspectable architecture — because with no transcript in the middle, there is nothing to evaluate, nothing to show you in a weekly report, no error record to practise from, and no way to measure whether your child is actually improving. Every step of our pipeline is observable, correctable, and parent-visible. For a children's product, we believe that is not a limitation. It is the point.
What Kiki is designed to do
- Hold a short, guided Hindi conversation (around ten minutes) on your child's current practice topic, pitched just above their level.
- Correct gently, the way a fluent relative would — by naturally using the right form in her reply, not by saying "wrong."
- Reward effort and speaking, never punish mistakes — a mistake never costs stars.
- Remember your child — their dog's name, the words they've earned, the errors they keep making — and build practice missions from their own mistakes.
- Report to you every week: what your child actually said, verbatim, and what Kiki will work on next.
What Kiki deliberately can't do
- Have an open-ended conversation. Kiki has no general chat mode; anything outside the lesson gets a warm redirect back to it.
- Browse the web, discuss the news, or bring outside content into the conversation.
- Contact your child. Kiki exists only inside a practice session that is started on your account, and never messages anyone.
- Show ads or pitch purchases to your child. There is nothing to buy inside a conversation.
- Connect your child with anyone. There is no chat with other users, no social features, no strangers.
Where it can go wrong — our known limitations
No AI system is perfect, and we would rather you hear the imperfections from us:
- Kiki can mishear. Accents, background noise, and soft little voices cause transcription mistakes. This is why every recording is kept and playable by you — if the transcript looks odd, you can hear exactly what your child said.
- The grammar evaluation is not infallible. Occasionally it misses an error, or flags something that was fine. Corrections are therefore delivered as gentle, reviewable notes — never as penalties.
- Kiki's own Hindi can occasionally be stilted. She speaks well, but like any AI, not perfectly. We review conversations continuously and tune her speech.
- Kiki is practice, not a teacher. She is the daily speaking layer — she does not replace a Hindi class, a school, or a grandparent. She exists to make those conversations possible.
What we don't do — stated plainly
We do not train our own frontier AI models; we build carefully on top of the best available ones. We do not yet run a formal adversarial "red-teaming" programme — the conversation surface is narrow by construction, but as we scale, formal adversarial testing is on our roadmap, along with a fairness review of how our speech recognition handles different accents and dialects. We would rather name what is ahead of us than pretend it is behind us.
For what data is collected and who can see it, read our Child Safety & Data Promise. For the formal legal version, see the Privacy Policy & Terms.
Last updated: July 2026 · Hindi Tutor is a product of Prayaas Impact Technologies. Questions about anything on this page: shubham@hindispeakingtutor.in