adam-s/car-diagnosis

Cardiag is a command-line and web tool that listens to a recording of a car and tries to identify what is wrong with it. You point it at an audio clip, and it returns a verdict about whether something sounds abnormal, which area of the car it appears to be coming from, and a ranked list of likely faulty parts. The tool is built as a triage aid, not a diagnosis replacement. The authors are clear about what it can and cannot do: on phone-quality recordings it achieves a 0.79 area under the ROC curve for detecting faults (versus a 0.50 score for random guessing). It gets the correct car zone in the top-3 results about 75 percent of the time. Crucially, when the audio is too ambiguous to call, it says "uncertain" rather than guessing. The README includes honest benchmark numbers and an explanation of a head that was cut for failing out-of-sample testing. The pipeline has several stages. First, it scrapes audio examples from YouTube or TikTok, or accepts clips you record yourself. Then a cleaning step strips out speech, music, and road noise to isolate the mechanical sounds. A pre-trained audio model called CLAP converts those sounds into a numerical representation. Finally, small linear classifiers trained on top of that representation produce the fault, zone, and part predictions. A pre-trained model ships with the repository, so you can run a diagnosis immediately after cloning without downloading additional data. The web app version lets you drop in a clip or paste a YouTube link and see the result, including an explanation of why the model returned that answer. A separate inspect command generates a visual and audible breakdown of what the pipeline extracted from your clip. The cleaning recipe and the honest training approach are noted as the main reusable contributions. The same method reaches 0.93 on cleaner engine audio datasets, suggesting the limitations come from the difficulty of phone recordings rather than the technique itself.