AIQ
AIQ
Everything a teacher needs to deliver this lesson — pick your grade's script below once you've read the background.
This lesson runs 5–15 minutes inside the app and is the audio half of World 2's opening pair — "How Computers See" covers images, this one covers sound. Both teach the same underlying idea from a different sense: a computer doesn't perceive anything the way a person does. It converts the raw signal into numbers, then an AI model finds patterns in those numbers. Here the raw signal is sound.
By the end of the lesson, a student should be able to:
Start from what sound actually is, because the rest of the lesson only makes sense once this is solid: sound is air vibrating. When you speak, your vocal cords make the air around your mouth vibrate in a pattern, and that vibration travels outward as a wave until it reaches an eardrum — or a microphone. A microphone is a small membrane that moves back and forth with the air pressure hitting it, and that physical movement gets converted into an electrical signal that rises and falls in the same pattern as the sound. At this point, nothing "smart" has happened yet — this is the same process a telephone or a tape recorder from decades ago used. A computer's next step is what makes this an AI lesson: it measures, or "samples," that electrical signal thousands of times every second — the lesson uses the real, standard number, 44,100 times per second — turning a continuous wave into a very long list of numbers. That list of numbers is the only thing the computer actually has to work with. It never "hears" a voice; it only ever sees numbers.
The next idea, the spectrogram, is the one worth spending the most classroom time on, because it's the closest thing this lesson has to a genuine "aha." A long list of raw amplitude numbers is hard for a model to find speech patterns in directly, so the numbers get transformed into a picture instead: a spectrogram is a chart with time running left to right, frequency (pitch, roughly) running bottom to top, and color or brightness showing how strong each frequency is at each moment. A sustained vowel sound shows up as a bright horizontal band; a sharp consonant shows up as a brief vertical smear across many frequencies. This is the sense in which "AI sees sound" — the audio problem gets converted into an image problem, and the same kind of pattern-recognition model used for photos can be pointed at this picture instead.
The wake-word detail matters for a reason beyond trivia: it's a real, deliberate engineering choice with a privacy angle worth surfacing, especially for the two older bands. If a phone sent every sound in the room to a server all the time, that would be both an enormous privacy problem and an enormous data cost. Instead, a small, low-power model runs constantly on the device itself, trained to recognize only one short phrase — "Hey Siri," "OK Google," "Alexa." That tiny model is not doing full speech recognition; it's doing one narrow, cheap task extremely well. Only once it fires does the device wake up a much larger, more capable model — frequently sent to the cloud because it needs more computing power than a phone has — to actually work out the words in the rest of the sentence. Bird-song ID apps like Merlin, sound-fingerprint apps like Shazam, and anomaly-detecting security cameras all use variations of the same underlying skill: a model trained on huge numbers of labeled sound examples (birdsong recordings, catalogued songs, recordings of breaking glass versus barking) to recognize a category of sound pattern it has seen before.
For the oldest band, it's fair — and accurate — to name the real current technology: modern speech recognition systems, including OpenAI's Whisper, are Transformer-based encoder-decoder models trained end-to-end on very large amounts of audio paired with text. That's a real generational shift from the statistical methods (like Hidden Markov Models built on hand-engineered features called MFCCs) that dominated speech recognition for decades before deep learning took over. Two more specific ideas are worth having ready if a curious student pushes further: CTC (Connectionist Temporal Classification) is one solution to a real technical problem — audio frames and text characters don't line up one-to-one, since a sound can stretch across many frames — and self-supervised methods like wav2vec learn useful audio representations without needing millions of hand-labeled transcripts, by training a model to predict a masked-out chunk of audio from the surrounding context, similar in spirit to how a language model learns by predicting missing words.
No prep and nothing to print. Each student (or pair, if devices are shared) needs a phone, tablet, or computer with a browser and the AIQ app loaded — the lesson's hook, its three learning scenes, and its quiz all run entirely on-device with no login and no actual audio recording required, since the content is taught through text, emoji, and short facts rather than playing real sound clips. If a student's own phone has a voice assistant enabled ("Hey Siri," "OK Google," or similar), it's a nice optional live demo during the hook — but it isn't required, and several phones triggering at once in a shared classroom will just create noise, so keep it to one volunteer's device if you use it at all.