AI Car Diagnostic Guide

The diagnostic engine works by converting your recording into a dual-panel spectrogram — a visual representation of frequency versus time — and matching it against a library of confirmed fault signatures. Different mechanical failures produce distinct acoustic fingerprints: rod knock generates periodic low-frequency impacts at crankshaft frequency, wheel bearing failure produces a speed-proportional hum with characteristic amplitude modulation, and timing chain rattle appears as a transient broadband spike in the first seconds of cold startup.

Context you provide — when the sound occurs, whether it is speed-dependent or RPM-dependent, whether it is present at idle — is applied as a set of hard gates before visual analysis. A sound that only occurs under braking cannot be rod knock; a sound that disappears above 20 mph cannot be a failing starter. These constraints meaningfully improve precision on ambiguous recordings.

Twelve of the 19 conditions achieve near-100% top-1 accuracy. The remaining seven are limited by the physics of audio — several noise sources produce acoustically similar signatures and require supplemental OBD-II data or physical inspection for definitive identification. Those accuracy ceilings are documented on each condition's page.