A recording of keyboard clicks can reveal which keys were pressed. In a 2023 study, researchers reported 95 percent classification accuracy from phone recordings and 93 percent from audio recorded through Zoom.
Those figures describe a trained model and specific experiments. They do not mean any microphone can recover 95 percent of every password from an arbitrary laptop.
Joshua Harrison, Ehsan Toreini and Maryam Mehrnezhad studied an acoustic side channel, information exposed by the sound of an operation. The computer does not intentionally send the typed character through its microphone. Physical differences in the clicks can still carry clues.
The experiment used a 2021 MacBook Pro and a nearby iPhone for one recording setup. A second dataset came from the laptop's microphone through a Zoom recording. The researchers collected examples of known keys before testing their classifier.
Their processing converted the recordings into visual representations of sound frequencies. A deep-learning model then learned to associate those patterns with particular keys.
The distinction between classifying keys and reading an entire message matters. Real text includes combinations, corrections and pauses. A laboratory accuracy for individual recorded keystrokes is not the same measure as reliable recovery of complete conversations.
Even with those limits, the experiment broadens the meaning of a data leak. Information can leave through ordinary sound rather than a software vulnerability or network packet. The keyboard has an acoustic signature.
