How Voice Assistants Work: A Middle School Lesson

A student raises a hand halfway through your grammar lesson: “Wait — how does Alexa even know what I’m saying? Like, actually know?” You look up. It is a fair question. A genuinely interesting one. And there is no chapter in the curriculum that answers it. The problem is not that you weren’t paying attention during a PD session on voice AI. The problem is that the lesson was never written. This post is that lesson — a concrete walk through how voice assistants process speech, why they mishear people, what they do with your audio, and a no-device role-play that makes the whole pipeline visible in one class period.

TL;DR: Voice assistants like Alexa, Siri, and Google Assistant follow a five-step pipeline: wake-word detection, automatic speech recognition (sound → text), natural language understanding (text → intent), response logic, and text-to-speech output. Every step is a separate AI subsystem. Understanding the pipeline helps students explain why voice assistants mishear them, what “always listening” actually means, and why accent and background noise cause failures — not glitches, but predictable, teachable limits. This post gives grades 6-8 teachers a how-voice-assistants-work lesson for middle school with a full unplugged activity, three standards crosswalks, and ready-made extension resources.

What happens when you say “Hey Siri”? The 5-step pipeline

When a student says “Hey Siri, what’s the weather?” five distinct subsystems fire in under two seconds. Each one can be named, explained, and role-played without any devices in the room. The pipeline is the whole lesson.

Step	What it does	Everyday example
1. Wake-word detection	The device listens for one specific sound pattern and ignores everything else	Alexa hears “Alexa” but not “Alexandra” or “elect her”
2. Automatic Speech Recognition (ASR)	Converts the audio waveform into a text transcript — sound → text	”Turn off the light” becomes the string of words the system will act on
3. Natural Language Understanding (NLU)	Parses the text to extract intent (what the user wants) and entities (which things)	“Turn off the light” → intent: control device; entity: light; action: off
4. Dialog / response logic	Looks up the answer, runs the action, or generates a reply	Checks the smart-home connection, finds “bedroom light,” sends the off command
5. Text-to-Speech (TTS)	Converts the text reply back into spoken audio	”OK, I turned off the bedroom light” — played through the speaker

Steps 1 and 2 map directly to AI4K12 Big Idea #1 (Perception) — the microphone is a sensor converting physical sound waves into digital data the system can process, exactly as a camera converts light into pixels. Steps 3 and 4 connect to AI4K12 Big Idea #4 (Natural Interaction) — the core promise of voice AI is that humans can communicate with machines in ordinary language rather than code.

How does speech recognition actually turn sound into words?

The short answer

ASR breaks the audio into tiny time slices (roughly 10-millisecond frames), identifies the probable phonemes — the smallest units of sound — in each slice, and then runs probability math to find the most-likely word sequence that fits those phonemes in order.

Think of it like autocomplete on a phone keyboard, but for sound. When you type “tomor,” the phone suggests “tomorrow” because that sequence of letters most often precedes a space in its training data. ASR does the same thing with sound patterns: given these phonemes in this language context, the most probable word is “weather,” not “whether.” The difference between a reliable voice assistant and a frustrating one is how large and how diverse that training set was.

This is the concept behind ISTE 1.1.d — understanding fundamental concepts of technology operations. Students who can explain that ASR is a probability model, not magic, are applying exactly the kind of foundational technology literacy that standard describes.

Why does your voice assistant mishear you?

The short answer

Every failure mode in ASR is a mismatch between the conditions the model was trained on and the conditions it now faces.

When you see that moment in class — a student’s voice assistant returning something absurd — that is not a broken product. That is a system hitting the edge of its training data. The mishearing is the lesson. Here is a table of the most teachable failure modes:

Why it fails	What happens	Classroom talking point
Accent and dialect bias	The model was trained mostly on majority-accent recordings; minority-accent speakers get lower accuracy	A Stanford-led study found the average word error rate was nearly double for Black speakers (0.35) versus white speakers (0.19) across five major ASR systems (Koenecke et al., 2020, PNAS)
Background noise / false wake	A nearby word triggers the wake-word detector; or noise masks phonemes	”Can you hand me the Lexa report?” accidentally wakes Alexa
Homonyms	”There,” “their,” and “they’re” sound identical; the system picks based on context probability	A student asking to “go to their page” might get “go to there page” in a transcript
Fast or quiet speech	Phoneme boundaries blur; the model guesses across ambiguous frames	Soft-spoken students get more errors as the microphone captures less clear signal

This is AI4K12 Big Idea #3 (Learning): the model is only as good as its training data. If the training data skewed toward certain voices, the system skews too. That is not an accident — it is a property of how machine learning works, and it is entirely within a grade-7 student’s comprehension to name it.

Is your voice assistant always listening?

The short answer

Before the wake word: the device processes audio locally and immediately discards it. After the wake word: audio is recorded and sent to company servers. After the response: the clip may be stored for quality review, depending on user settings.

Moment	What the device does	What data leaves the device
Before the wake word	Runs a tiny on-device wake-word detector; no recording	Nothing
During the command (after wake word)	Records the full utterance and streams it to cloud servers	Your voice audio + device ID + timestamp
After the response	Cloud server processes, logs, and (by default on most platforms) stores the clip	Stored clip + associated metadata, subject to platform privacy settings

This is honest, not alarming. Students who understand this table can make real decisions: turn off voice history storage in account settings, choose a voice assistant with on-device processing, or simply know what they are agreeing to when they set one up at home. That is ISTE 1.2.d in action — managing personal data to maintain digital privacy and security.

A no-device unplugged activity to teach the pipeline

This is the voice assistant unplugged activity that lets every student physically become part of the processing chain.

Setup (5 minutes). Assign five roles — one per pipeline step. Give each student a role card with their step name and one job description sentence. The teacher acts as the “user” — the voice input source.

Round structure (20-25 minutes). The teacher says a short command aloud: “Hey class assistant — set a timer for three minutes.” The pipeline runs in sequence:

1. Wake-word student listens for the trigger phrase and signals thumbs-up when heard — otherwise stays silent.
2. ASR student writes the spoken words on a notecard as a text transcript and passes it forward.
3. NLU student circles the intent and entities on the notecard (intent: timer; duration: 3 minutes) and passes it forward.
4. Dialog student decides the correct response action and writes the reply text.
5. TTS student reads the response aloud in a deliberate, even voice.

Run three rounds. In the second round, speak in an exaggerated accent and watch where the ASR student struggles to transcribe accurately. In the third round, add background noise — someone tapping a desk — and watch the wake-word student trigger at the wrong moment. The breakdown is the demonstration.

Debrief question (5-10 minutes). “Which step caused the most errors in round two — and whose job would it be to fix it?” Students typically name ASR. Follow up: “If you were building a better ASR system, what would you change about the training data?” A typical class finishes all three rounds plus debrief in 30-40 minutes. This maps to CCSS.ELA-LITERACY.SL.7.1 — engaging in collaborative discussions with diverse partners, building on others’ ideas, and expressing their own clearly.

For more no-device formats, the Unplugged AI Activities — 10 No-Tech Lessons collection includes card sorts and role-plays across multiple AI concepts for grades 6-8. The broader question of where voice assistants fit in the AI taxonomy — as narrow AI — is in the Types of AI Deep-Dive Lesson.

How this lesson maps to standards

Activity step	Standard code	What it shows
Microphone as sensor; pipeline overview	AI4K12 Big Idea #1 — Perception	AI systems use sensors to gather data from the physical world
Accent and training-data failure modes	AI4K12 Big Idea #3 — Learning	Models are shaped by — and limited by — their training data
Natural language as human-AI interface	AI4K12 Big Idea #4 — Natural Interaction	Humans use everyday language to communicate with AI systems
ASR probability model explanation	ISTE 1.1.d	Students understand fundamental concepts of technology operations
Privacy table and settings discussion	ISTE 1.2.d	Students manage personal data to maintain digital privacy
Pipeline role-play and debrief	CCSS.ELA-LITERACY.SL.7.1	Collaborative discussion with diverse partners, building on ideas

Standards source: AI4K12 Initiative; ISTE Student Standards. ISTE is a registered trademark of the International Society for Technology in Education. These resources are not affiliated with or endorsed by ISTE.

Ready-made lessons and where to go next

Nobody handed you this curriculum — and that is exactly the gap these lessons exist to close. You were not given the tools for teaching how voice assistants work; the materials simply did not exist in most buildings until recently.

The unplugged pipeline activity in this post pairs directly with the AI Everyday Life Lesson Bundle — a three-lesson set that puts voice assistants, recommendation engines, and image recognition side by side so students see how narrow AI shows up across their daily environment. For the broadest concept coverage, the How AI Works MEGA — 16 Concept Lessons treats voice assistants as one of sixteen AI concepts in a full-year sequence for grades 6-12.

To extend the language-processing angle of this lesson, see how the language model side works — what happens after NLU when a system generates its reply. For more no-device formats, the unplugged machine learning activities collection has additional role-play options across different AI mechanisms. If you want to see how voice assistants relate to more autonomous AI systems, the AI agents post draws the line clearly. And if you want to start with something free before committing to a unit, the starter resources page has printable options.

This post was drafted with AI assistance and human-finalized.