A side channel is information that leaks from how a system runs rather than from the data it processes. Timing, power draw, electromagnetic emissions, and sound all qualify. The acoustic version is unsettling because the sensor is already everywhere: the microphone in your laptop, your phone on the desk, the smart speaker on the shelf, the other people on your video call.
Why Keys Sound Different
A keyboard is a grid of physical switches under a rigid plate. When you press a key, the sound that reaches a microphone depends on where on the plate the impact happened, how the case resonates, and the tiny differences in how each switch actuates. A key at the edge sounds subtly unlike a key in the center. To a human ear the clicks blur together. To signal analysis they carry a fingerprint.
There is a second, richer signal in the timing. People do not type at a constant rate. The gap between two keystrokes depends on the physical distance your fingers travel and on how common that letter pair is in the language. Those inter-key intervals narrow the possibilities enormously, the same way a partly filled crossword constrains the remaining answers.
An acoustic attack combines the per-key sound signature with the timing between keystrokes. Neither is decisive alone. Together, backed by a language model that knows which words and letter sequences are plausible, they turn a stream of clicks into readable text.
From Research Curiosity to Practical Attack
This is not new in principle. As far back as 2004, researchers at IBM showed that keystrokes could be distinguished by sound, and follow-up work added statistical language modeling to improve the guesses. What was missing then was the ease of collection and the strength of the classifier.
Both have changed. Deep learning turned the messy job of matching a sound to a key into a well-understood classification problem, and the microphones are better and more numerous. In a widely reported 2023 study, researchers trained a neural network on the sound of a laptop keyboard and recovered keystrokes with high accuracy from a nearby phone recording, and with only slightly lower accuracy from audio captured over a video call. The exact numbers depend heavily on the setup, but the direction is clear: the barrier to this attack has dropped from specialist to feasible.
The microphone does not need to be compromised in any exotic way. It can be the one you deliberately turned on to join a meeting.
What Makes an Attack Easier or Harder
| Factor | Effect on the attacker |
|---|---|
| Same keyboard for training | If the model is trained on recordings of your specific keyboard, accuracy climbs. A generic model trained on a different device does worse. |
| Typing style | Steady touch typing produces cleaner, more separable keystrokes. Hunt-and-peck typing and heavy use of correction keys add noise. |
| Background noise | Music, conversation, and room noise degrade the signal. A quiet room is the attacker's friend. |
| Predictable content | Prose is easy because a language model fills gaps. A random password has no linguistic structure to lean on, which cuts both ways. |
That last row deserves a caveat. A long random password is harder to reconstruct in full, but a password is often a short, high-value target, and even a partial recovery that narrows it to a handful of candidates can be enough when combined with an online guessing attempt. Do not assume randomness alone defeats the attack.
What Actually Helps
There is no single fix, and some intuitive ideas do little. Typing more randomly is impractical and unreliable. The defenses that hold up are a mix of behavior and design:
- Do not type secrets during a call. The single highest-value habit. If you must enter a password or passphrase while on a video conference, mute your microphone first.
- Prefer a password manager and autofill over typing credentials by hand, so the most sensitive strings are never keyed in audibly. See password manager security.
- Use a second factor so that a recovered password alone is not enough. A hardware security key is not something an acoustic attack can capture.
- Move to passwordless where you can. A passkey is never typed, so there is no keystroke to overhear.
- Mind your microphone hygiene. Know what is listening. A phone face-down and muted, a laptop mic disabled when not in use, and awareness of smart speakers all shrink the sensor surface.
How Much Should You Worry
Threat modeling matters here, because the honest answer is not the same for everyone. This is a targeted, proximity-dependent attack. It needs a microphone near you, a reasonably quiet environment, and in its strongest form some data about your specific keyboard. It does not scale the way a phishing campaign or a credential-stuffing run does. Nobody is acoustically transcribing millions of people at once.
For most people, most of the time, it sits low on the list of realistic risks, well below reused passwords and unpatched devices. It climbs the list if your work makes you a specific target: a journalist meeting a source in a shared workspace, a lawyer typing privileged notes on a call with an adversary present, an executive on an earnings call, anyone whose keyboard is regularly within earshot of a device or a person they cannot fully trust. The people who need the strongest operational security are exactly the ones for whom a microphone in the room is a genuine exposure, not a curiosity.
The reason to know about it even if you are not a target is that the countermeasures cost almost nothing. Muting before you type a password and moving to passkeys are good habits regardless of who is listening, and they close this channel as a side effect.
The Broader Point
Acoustic keystroke recovery belongs with motion-sensor side channels and electromagnetic emanation attacks in a category that end-to-end encryption cannot touch: they capture information at the endpoint, before encryption begins or after decryption ends. It is worth distinguishing this from keystroke dynamics, which fingerprints who is typing by rhythm rather than recovering what they type.
Strong cryptography is necessary and it is not sufficient. Haven protects the message in transit and at rest with client-side encryption, which is the hard part to get right and the part most services get wrong. The endpoint, the device, the room, the microphone, remains a shared responsibility. Knowing that the walls can listen is the first step to not saying the sensitive thing in front of them.