Teaching "How Computers Hear" to Builder mode (ages 8–10)

Hook & Warm-Up

Open with the app's own line, read with real curiosity rather than as a script:

"'Hey Siri!' 'OK Google!' How does a machine understand your voice? It doesn't speak any language — it reads sound waves like sheet music!"

Ask the class: "If a computer doesn't actually understand English or Filipino or any language at all, how do you think it figures out what you just said?" Take two or three guesses and write them on the board without judging them yet — you'll come back to these at the end of the lesson. Then say: "Today we're becoming Sound Detectives 🎤 and finding out the real answer — and it starts with something surprising: your voice becomes a picture before a computer ever 'reads' it."

Add a quick show of hands before opening the app: "Who here has a smart speaker, a phone, or a tablet at home that responds to a spoken wake word?" Then ask a follow-up that most students won't be able to answer yet, which is the point: "Does it understand every language in the world, or just the ones it was specifically taught?" Hold onto that question — the wake-word part of today's lesson answers it directly.

Main Activity

Open the lesson on the shared screen and work through its three scenes together, pausing after each item for a quick check rather than reading straight through. The three scenes build on each other in a specific order — first how sound becomes something a computer can even look at, then how that leads to actual speech recognition, then a tour of other things built the same way — so resist the urge to skip ahead even if a scene looks familiar.

Scene 1 — Sound Waves 🌊. Walk through the four facts in order: sound is really just air vibrating (have students tap a desk and feel the vibration travel through the wood); a microphone turns those vibrations into an electrical signal, which a computer can graph as a wiggly line; a computer then samples that signal — measures it — 44,100 times every single second, to capture every tiny detail; and finally, AI converts that stream of numbers into a spectrogram, a picture that shows which pitches are happening at which moments. Draw a very rough spectrogram on the board — a few tall, dark smudges — and say: "This is genuinely what the computer is 'looking at.' It's never dealing with an actual sound, only ever this kind of picture." Ask the class why they think 44,100 is a big number to check every second — most will land on "so it doesn't miss anything," which is close enough; you can add that this exact number is the same one used for regular music CDs, so it's a real, standard number, not something the app made up.

Scene 2 — Speech Recognition 🗣️. Explain wake words first, since it connects directly to the hook: "'Hey Siri' works because a small AI is running right there on your phone, all the time, only trained to notice those two words — nothing else." This is also the moment to circle back to your opening question about whether a device understands every language: it doesn't — the wake-word model, and everything downstream of it, is trained only on the languages its maker chose to support, which is exactly why a voice assistant can fail badly on a name or word from a language it wasn't trained on. Then explain how a full sentence gets recognized: AI breaks speech into tiny pieces called phonemes (write "H-EH-L-OW" on the board for "Hello" and sound it out with the class), then matches those sound pieces to words it already knows — "that's the same trick behind dictation apps and the captions on a video." Mention Google Translate as the most impressive combination: it's really two (or three, counting the voice that speaks the answer back) AI systems running back to back, one for hearing your language and one for speaking a different one — a chain, not a single system that magically does everything at once.

Scene 3 — Sound AI 🎵. Cover all four examples and ask a quick "what pattern is it matching?" question after each: Shazam builds a sound fingerprint and searches it against millions of songs (matching a fingerprint); the Merlin app identifies birds from thousands of training recordings (matching a species' typical call); Nest cameras tell a dog bark apart from breaking glass (matching two very different sound patterns); AirPods listen to outside noise and generate an opposite wave to cancel it (matching, then countering, a pattern in real time). It's worth pointing out to students that these four examples aren't all doing the exact same kind of job — Shazam and the bird app are both trying to answer "which one of these known things is this?", while the noise-canceling headphones are doing something closer to "generate the opposite of whatever I just heard" in real time. Both count as sound AI, but they're solving different problems.

Close with the app's own summary: "Computers turn sound into numbers — sampling 44,100 times per second! AI then finds patterns in those numbers to recognize speech, music, and sounds." Then let students complete the in-app practice round, sorting items into AI and not-AI on their own or in pairs.

While students work through the practice round, circulate and ask individual students to explain one of their answers out loud — not to catch anyone out, but because saying "it's AI because it learns a pattern from lots of examples" in their own words is a stronger sign of real understanding than getting the right answer by process of elimination. If a few students consistently confuse "electronic" with "AI" (a common mix-up, since a wired telephone and a doorbell are both electronic), that's worth a thirty-second whole-class pause to reinforce the distinction before moving to the quiz.

Discussion

• "We wrote down some guesses at the start about how a phone 'understands' you. Which guesses were closest to the real answer, now that we know it's sampling and spectrograms?"
• "Why do you think 'Hey Siri' runs on your phone itself, instead of always sending everything you say to a big computer far away?" (Guide toward speed and privacy, even if students only get partway there on their own.)
• "A wired telephone from a long time ago also carries your voice as an electrical signal — so why isn't a wired telephone AI?" (Because nothing in it learns patterns or recognizes anything — it just carries the signal through.)
• "Can you think of a sound at home that would be genuinely hard for an AI to recognize? Why might it be hard?"
• "If a voice assistant is only trained on certain languages, what do you think happens when it hears a word it's never been trained on — a Filipino nickname, a local dish, or a place name?" (Guide toward: it will likely mishear it and guess the closest sound-alike word it does know, rather than genuinely recognizing something outside its training.)
• "Noise-canceling headphones make an opposite sound wave to cancel outside noise. Why do you think this only works well for steady sounds like an airplane engine's hum, and not so well for a sudden sound like someone dropping a book?" (Guide toward: canceling requires quickly figuring out the shape of the incoming wave and generating its opposite — a steady, repeating sound is far easier to predict than a sudden, one-time noise.)

Quiz Walkthrough

Wrap-Up & Extension

Close with: "Next time you say 'Hey Siri' or ask Shazam what song is playing, you'll know what's really happening behind the scenes — your voice becomes a picture, and an AI trained on millions of other pictures finds the closest match."

Extension activity — "Build a Fingerprint": In pairs, have students pick three very different sounds they could make with objects in the room (a pencil tapping, a page flipping, a chair scraping) and describe each one in writing using only shape words — "short and sharp," "long and low," "fast and bumpy" — no naming the object. Swap descriptions with another pair and see if they can guess which sound matches which description. This is a hands-on stand-in for what a sound fingerprint actually captures (a distinctive shape, not the object itself), and stretches the 10–15 minute core lesson into a full 30–35 minute period. If a group finishes early, challenge them to make two of their three sounds as similar as possible on purpose (two different but similar taps, for example) and see if the other pair can still tell them apart from the written description alone — that's a small, concrete taste of why sound AI sometimes makes mistakes on genuinely similar sounds, the same way a person might mishear two close-sounding words.