What is kbd-audio?

kbd-audio is a collection of command-line and graphical tools that demonstrate a privacy attack called acoustic keyboard eavesdropping: figuring out what was typed on a keyboard by listening to the sounds the key presses make through a microphone. The most prominent tool is called Keytap. In its first version, Keytap analyzes live microphone audio and attempts to identify which keys were pressed, based on the distinct sound profile each key makes. It works best with some training data first: recordings of the target keyboard through the same microphone, so the system learns what each key sounds like on that particular setup. A second version, Keytap2, removes the need for training data by combining audio analysis with statistical knowledge about the English language. It looks at how often certain letters appear and how letters tend to follow one another, then uses that structure to narrow down what was likely typed without any prior recording of the keyboard. The third version, Keytap3, builds on that approach with improved algorithms and more complete statistics about letter combinations. It is described as fully automated, meaning no manual adjustment is needed during the text recovery process. Online demos for all three versions are available, including a challenge page where you can test whether a recording of your own typing is recoverable. Beyond the Keytap tools, the repository includes utilities for recording audio during typing sessions, playing back recordings, visualizing audio waveforms, and detecting individual key presses. The tools are written in C++ using SDL2 for audio capture and window rendering. They can be compiled from source on Linux, macOS, and Windows. The project is from Georgi Gerganov, who is also known for other open-source projects in audio and machine learning.