🔊 Machines That Hear · Day 13

Stop the Echo

On a video call, the other person's voice comes out of your speaker… and leaks straight back into your mic. Without a fix, they'd hear themselves as an annoying echo! Phones cancel it by a clever trick: they already know what came out of the speaker, so they subtract it back out.

↓ cancel the echo
🏷️ The pro wordThis is Acoustic Echo Cancellation (AEC). The mic's copy of the speaker sound is the echo path, and the smart filter that learns to remove it is an adaptive filter. Every phone, laptop and smart speaker runs AEC so calls don't echo.
🎚️ Lab · The Echo Canceller

Subtract the echo back out

🟦 Far voice = what came out of your speaker. 🟥 Your mic = it hears that voice bounce back (delayed & quieter) — the echo. The AEC guesses the echo's delay and volume, and subtracts it → 🟩 clean. Tune the two sliders until the echo meter hits zero!

0

estimated echo delay (how long the bounce takes)

0.0

estimated echo volume

Echo still leaking
100%
🎯 Goal: get the echo below 8% — that's a cancelled echo, and a happy phone call!
💡 The Real Challenge

Why it's tricky in real life

You just matched the echo by hand. A real AEC has to do this automatically and continuously, because the echo changes when you move your phone, change volume, or the room changes. That's why it uses an adaptive filter that keeps re-learning the echo path many times a second.

🗣️ Extra-hard case: double-talk — when both people speak at once. The AEC must cancel the echo but not your real voice. Getting that right is a big deal in call software!
🚀 Beyond the Basics

How the pros really do it

You matched the echo's delay & volume by hand — great for seeing the idea. But the real echo keeps changing, so pros make it automatic:

🟢

Hand-matched (what you built)

✅ Shows exactly how AEC works. ⚠️ Only cancels if nothing moves or changes.

🔁

Adaptive filter (auto-tuning)

Algorithms that re-learn the echo path many times a second as you move the phone. ✅ Handles real, changing rooms — every phone uses this. ⚠️ Math-heavy.

🧠

Deep-learning AEC (AI)

AI handles the hardest cases: speaker distortion and double-talk (both people speaking). ✅ Cleanest calls. ⚠️ Needs a model + compute.

🏷️ Pro names to look upadaptive filters (NLMS / RLS) · WebRTC AEC (open-source, in your browser) · deep-learning AEC.

🎉 Day 13 Complete!

You performed Acoustic Echo Cancellation — the trick that keeps every call from echoing, by subtracting the known speaker sound back out of the mic.

Day 14: Clean the Voice →
Tomorrow: echoes gone — but what about the fan, traffic, and hiss? Next: noise suppression. 🌬️
AI Adventure · Machines That Hear · Day 13 — made for young inventors 🚀