This blog is written by Gokul Kumar from Madhi Foundation.
Classrooms are loud places. A teacher reading aloud, thirty children chorusing back, a child muttering through a subtraction problem, a coach giving feedback in the corridor after class. Almost none of that ends up in our data.
Over the past few months we’ve been running an experiment: does conversational voice AI actually work in a government classroom in Tamil Nadu? We’re using ElevenLabs — which gives us speech recognition, an agent that can hold a real back-and-forth, and text-to-speech natural enough that a teacher doesn’t wince and a seven-year-old doesn’t lose interest — and testing it in two places that couldn’t be more different: teacher reflection in VallamAI, and student support in Enjoy English.
This is an early write-up. We don’t have results yet. We have a working hypothesis. Experiment 1: Teacher reflection in VallamAI
VallamAI is live across 7 districts in Tamil Nadu, with over 2,100 teachers and 150 instructional coaches. One part of it is the Classroom Journal — a place for teachers to record what happened in a lesson, which children struggled, what they’d do differently.
So we’re testing it as a conversation. The teacher finishes a lesson, taps once, and talks — in Tamil, English, or the mix of both that’s how people actually speak.
What we wanted to find out was whether a voice agent gets better information than a form. Early signal says yes, mostly because it asks the second question. A teacher says, “Today’s lesson was a bit difficult.” A form records that and stops. The agent asks whether it was the whole class or a few children, and whether they were stuck on the concept or the language of the problem. Ninety seconds later we have something we can act on. The original sentence was a mood; this is a diagnosis.
The reflection then gets transcribed and mapped to things VallamAI already tracks — concepts covered, misconceptions surfaced, children to flag for remediation — which feeds the next lesson plan and helps coaches decide which school to visit.
Experiment 2: Student support in Enjoy English
Enjoy English is our mobile English learning app for Grades 1–5. It runs into a trap every early-grades app hits: the children who need the most support are, by definition, the ones who can’t read your on-screen instructions. A Grade 1 child still learning to decode cannot read “Tap the word that rhymes with cat.” So they guess, the app records a failure, and the child learns this is a place where they’re wrong a lot.
We’re testing voice as the way out of that. Instructions, hints, and encouragement spoken aloud, in a consistent voice children recognise as a character rather than a machine. And beyond narration, a support buddy a child can actually talk to — “I don’t understand this” — that responds without giving the answer away.
That last constraint has been the hardest part, and it’s pedagogical, not technical. General-purpose AI assistants are trained to be maximally helpful, which for a struggling child means handing over the solution. We’ve had to work hard to make the buddy scaffold instead: simplify, break it into a smaller step, ask what the child has already tried.
Voice also lets us assess things you can’t assess by tapping — reading fluency, phonics, letter-sound production. Teachers already lean this way; roughly 70% in our VallamAI deployment choose speech-based assessment over paper.
What we’re learning
We’re early, so this list is honest rather than flattering.
Student accents are our biggest open problem. Speech models are overwhelmingly trained on adult voices, and children’s speech is genuinely harder — higher pitch, less stable articulation. On top of that, our children speak English shaped by Tamil phonology, and it varies across regions. A child who reads a word correctly in their own accent should never be marked wrong. Getting this right is the difference between a useful assessment and a discouraging one, and we’re not there yet.
Teachers need probing that adapts. Right now the agent works from a mostly fixed set of questions. What would actually be valuable is follow-up shaped by that teacher’s own history — the child they flagged last week, a pattern across three lessons. That’s the next thing we’re building.
Trust is fragile. Reflection has to stay a support tool, never a surveillance one. The moment teachers suspect their journal is being read as an evaluation, they’ll stop saying anything true and the whole thing collapses.
Latency and bandwidth are real constraints. For a teacher, a one-second pause is thinking time. For a seven-year-old, two seconds is a broken toy. Meanwhile government schools don’t
have reliable connectivity, so we’re caching frequently-used audio on the device and letting live conversation degrade to record-and-sync rather than fail.
Classrooms have always been voice-first. We’re testing whether our software can finally agree.
