Teaching a Machine to Listen Like a Teacher: Our Children's Voice Data Journey
· Taimoor Haider · 7 min read
Over many months, our team sat with children of different ages and backgrounds and recorded their voices, one letter and one word at a time. The result is a large, carefully reviewed collection of children’s speech. This is the story of how that dataset came to be, and why it was one of the hardest things we have built.
At Makharij, our goal is simple to say and difficult to do: help children learn to pronounce every Arabic letter from its correct point of articulation, with feedback as patient and precise as a good teacher. To do that, our AI has to genuinely understand how children sound. Not adults speaking slowly. Not professional reciters. Children, with all their energy, hesitation and charm.
The data that didn’t exist
When we started, we assumed we could build on existing speech datasets. We were wrong. Children’s speech data is scarce everywhere, and children reciting Arabic letters with attention to makharij is almost nonexistent. What little exists is rarely recorded with the care that pronunciation assessment demands.
Adult recordings, which are far easier to find, turned out to be the wrong foundation. A child’s voice is not a smaller version of an adult’s:
- Higher pitch and different resonance. A shorter vocal tract shifts the acoustic patterns a model relies on to tell one sound from another.
- Still-developing articulation. Young children are learning to control the tongue, lips and throat, so their “near misses” look nothing like an adult’s mistakes.
- Different rhythm and breath. Children pause, stretch, rush and restart in ways adult speakers rarely do.
A model trained mostly on adults hears a child and gets confused, or worse, marks a correct letter as wrong. For a learning product, that erodes trust instantly. So we made a decision that shaped everything that followed: we would build our own children’s voice dataset from the ground up.
Finding the right microphone was its own project
We thought choosing a microphone would take a week. It took far longer.
The differences between many Arabic letters live in subtle details: a breath of air, a slight shift deep in the throat, the faint buzz of the nose. A microphone that sounds fine for a phone call can quietly erase exactly the details we needed to capture.
We tested a wide range of options and kept running into trade-offs:
- Some captured beautiful detail but picked up every chair scrape and ceiling fan in the room.
- Some rejected noise well but flattened the very sounds we cared about most.
- Some performed well in a quiet studio and fell apart in a real classroom.
- And children don’t sit still. Distance and angle to the mic changed from one word to the next.
Placement mattered as much as the device itself. Too close, and bursts of air distorted the recording. Too far, and the room took over. We went through many rounds of trial recordings, listening back carefully, before we settled on a setup and a recording routine that gave us consistent, clean audio across very different children and rooms.
We won’t share the exact setup we landed on. That knowledge was hard-earned, and it is part of what makes our models work. What we can say is this: if you are building speech AI for children, budget far more time for your recording setup than you think you need.
Recording where children actually are
A soundproof studio would have made our lives easier. It would also have produced a model that only works in a soundproof studio.
Children will use Awwal at home, in classrooms and in Quran learning centres. So that is where we went. We recorded in a mix of real-world spaces, each with its own challenges: echoing rooms, traffic outside, siblings in the next room, the call to prayer, generators, fans, and the occasional power cut halfway through a session.
We had to strike a careful balance. The audio needed to be clean enough to judge pronunciation accurately, yet realistic enough that the model would hold up in the noisy, everyday places where children actually learn. Finding that balance meant constantly adjusting: rearranging rooms, rescheduling around noisy hours, and sometimes simply waiting for silence before pressing record again.
A recording app built for children
Early on, we realised that a microphone and a laptop alone would not give us reliable recordings. So we built our own recording app with our full syllabus inside it. Before each session, the app checks that the microphone is connected and working, so no session is lost to a loose cable or a silent mic. While recording, it shows the noise level in the room and the child’s voice level. That way we could spot a noisy moment or a voice that was too soft or too loud before it spoiled a take. The child sees just one word on the screen at a time, with clear visual cues for when to start speaking and when to stop. This kept young learners focused and made each recording clean and easy to work with later. By handling these details automatically, the app greatly reduced human error and let our team focus on what mattered most: keeping the children comfortable and encouraged.
Every child sounds different
A five-year-old and a twelve-year-old may as well be speaking through different instruments. Their voices, their control over each letter, and even their attention spans are worlds apart. A model that only hears one age group will fail the others.
So we deliberately recruited children across a wide range of ages, and made sure both boys and girls were well represented throughout. Voices change in different ways for boys and girls as they grow, and we wanted the model to be fair to every child who uses it, not just the ones who happen to sound like the majority of the data.
We also looked for children at different stages of their learning: some just beginning to recognise the letters, others already reciting confidently. Beginners gave us the natural mistakes a learning tool must catch. Confident readers gave us strong examples of what correct pronunciation sounds like in a child’s voice. Both were essential.
One word, many takes
If there is one thing this project taught us, it is patience.
Children make mistakes, and that is the whole point of learning. But it meant that a single word could take many attempts before we had the recording we needed. A child would mix up two similar letters, giggle halfway through, look at a parent for reassurance, lose focus, or say it perfectly the moment we weren’t ready. Then we would smile, encourage them, and try again.
And the mistakes were never wasted. Real errors, made naturally by real children, are some of the most valuable data we have. They teach the model what a child’s mistake actually sounds like, so it can guide them gently toward the correct makhraj.
Keeping children engaged through a session became a skill of its own. We learned to keep sessions short, build in breaks, turn the process into something closer to a game, and celebrate small wins. A happy, relaxed child produces natural speech. A tired or anxious child doesn’t. Behind every hour in our dataset are many more hours of setup, encouragement, re-recording and careful review.
Trust comes first
Working with children’s voices is a responsibility we take seriously. Every recording was made with the informed consent of parents or guardians, and children were free to stop at any time. A parent, guardian or teacher was always close by.
We handle this data with care. Recordings are stored securely, access is restricted to the people who need it, and the data is used only to build and improve Makharij. The families who trusted us with their children’s voices made this work possible, and protecting that trust is non-negotiable.
Where we are now
Today, we have one of the very few datasets of its kind: a substantial collection of carefully recorded and reviewed audio from boys and girls across a wide range of ages. It is the foundation on which Makharij learns to listen to children the way a dedicated teacher would: closely, patiently and fairly.
This journey was slow, demanding and often humbling. There were no shortcuts, and we are glad we didn’t look for any. The difficulty is exactly what gives this work its value.
To every child who sat with us and tried again, to every parent who trusted us, and to every teacher and centre that opened its doors: thank you. Your voices are helping the next generation learn to recite with confidence.
We are continuing to grow our dataset. If you are a parent, teacher or Quran learning centre interested in taking part, we would love to hear from you at info@makharij.ai.