A research team from Cornell University in the United States has published a detailed explanation of how artificial intelligence can learn to capture information from the sound of keyboard presses. The AI system was able to predict typed text with up to 95% accuracy by analyzing keystroke sounds. When the model was trained through Zoom audio, the accuracy dropped slightly to 93%.

Achieving such high accuracy requires device-specific training; it cannot be done with a random keyboard. This can be performed locally using a microphone or remotely over a network using Zoom or a similar audio-capturing application.

In the project, the AI was trained on an Apple MacBook Pro. After just 25 keystrokes, the model learned to recognize the sounds of 36 different keys. Recognition accuracy can be significantly reduced by using a touch keyboard, introducing background noise, or simply changing your typing technique.

AI most easily recognizes input from mechanical keyboards, but it also works with other types — even membrane keyboards can produce enough acoustic variation for successful detection.