We counted the voices
in the room.
There was usually one.
Balochistan Winter School what 2,265 recordings heard.
6% of its population
Lowest literacy in Pakistan
Security checkpoints to reach schools
Internet blackouts routine
A winter morning in Quetta. Security checkpoints. No heating. Seventeen schools.
This is what the AI found when it listened to every lesson.
In January 2026, Taleemabad ran a Winter School across 17 Balochistan schools. Teachers recorded every lesson. An AI analysed every second of audio who spoke, how long, how often a student generated a question.
What the recordings found was not what we went in looking for. This is the raw signal, before the Georgetown RCT results land in May.
I was sitting in a classroom deep inside the Hazara community. You pass through security checks to get there. Freezing winter morning the heater wasn't working. I had layers on, a warm jacket, and I was still cold.
I quietly sat at the back, trying not to disturb. The teacher was teaching with high energy. Kids equally energetic moving around, talking on top of each other. Small stones on their desks. They were learning place values.
One of the most impoverished places on earth, and I heard children say for the first time that they love maths.
After class, the teacher uploaded his audio recording. The tool processed it and gave him feedback. I asked if he agreed. He said: "I can't believe how accurate it is."
I can't believe how accurate it is.
Teacher, Balochistan Winter School, January 2026
How AI Coaching Works
Most pilots drop off by week 3. This one didn't.
Weekly AI observations Balochistan Winter School
Weeks 5–6 were assessment phase lower observation count expected.
of lessons. Negligible student voice recorded.
Barely a question. Barely an answer. Student talk was measured at 7% of class time across 1,773 lessons. This is not a teacher failure — it is what happens in overcrowded classrooms where one person must manage 50 children and a full curriculum. The system was never designed for student voice. We just finally had a number on it.
Student questions Indicator D6
% of lesson time with student-generated questions
Student questions grew every active week. By week 3, a boy in a place values lesson raised his hand. His teacher had been coached to pause after asking a question to hold the silence. He did. The boy filled it. The AI had been asking for exactly that pause. 4.6% of lesson time with student questions may seem small. In a classroom where the baseline was zero, it is the first crack in the wall.
Learning Gains
This is what the teacher saw after the place values lesson.
You asked 8 open-ended questions
Try wait time after questions
Connect place value to everyday money
Every observation generates specific, actionable feedback. Not scores sentences. The kind a good coach would say.
2,265 observations processed
across 6 weeks · 17 schools
Teacher adoption faster than expected. 2,265 observations. Grew every active week. Teachers who'd never used a smartphone app were uploading consistently by week 3.
Real-life relevance (Indicator B8) 44% only. Hardest skill in the framework. The gap between content delivery and connecting learning to life remains the critical unlock for student voice.
Working in similar contexts?
We've seen the one voice. We want to know if you have too.
If you're working in conflict-affected or low-connectivity contexts and want to know what the signal looked like week by week reply to this email. We're building a small network of practitioners doing similar work.