July 4, 2026
Check engine… but make it gossip
Show HN: Classify mechanical faults using Contrastive Language-Audio Pretraining
This app wants to hear your car cough — and the crowd is already playing backseat mechanic
TLDR: A new demo can listen to car sounds and suggest what might be wrong, but it’s built to be careful and say “uncertain” when the audio is too messy. The comment section’s big reaction was classic internet skepticism: neat idea, but some already want old-school failure logic mixed in before trusting a phone recording over a mechanic’s ear.
A new Hacker News demo called cardiag is trying to do something delightfully chaotic: listen to your car’s weird noises and guess what might be wrong. The creator is very clear that this is not a magic mechanic in your pocket. It’s more like a cautious helper that says, “something sounds off,” points to a rough area of the car, and gives a shortlist of possible culprits — or admits it’s uncertain instead of making stuff up. That honesty got major “respect” energy, because internet demos usually promise the moon and then quietly explode on contact with reality.
But the real action was in the comments, where the community instantly shifted into armchair-diagnosis mode. The standout reaction came from yerbymatey, who dropped a research paper like a seasoned pit-crew plot twist, suggesting older-school fault-tree logic could help trace failures from symptoms. Translation for non-car nerds: not everyone thinks “let the sound model decide” is enough, and some want a more detective-style system layered on top.
That created the thread’s main vibe: cool idea, but can it survive the messiness of real life? The project scrapes noisy clips from YouTube and TikTok, strips out voices and music, and works from phone audio — which commenters clearly understand is a brutally messy source. The biggest joke hiding under the surface: people are already imagining their car being diagnosed between wind noise, bad roads, and someone yelling “what’s that clunk?” In other words, the community seems intrigued, cautiously impressed, and fully ready to second-guess the robot mechanic from the passenger seat.
Key Points
- •The project, cardiag, uses an audio cleaning pipeline, frozen CLAP embeddings, and linear classification heads to triage possible car faults from recordings.
- •The article presents the system as a proof of concept and calibrated triage aid, not a definitive diagnostic tool, and it can return UNCERTAIN when evidence is insufficient.
- •Reported leakage-safe evaluation over 1,031 video groups shows AUROC 0.79 for fault vs. normal detection, about 75% top-3 accuracy for six car zones, and about 45–65% top-3 accuracy for 12+ part families.
- •The same method is reported to reach 0.93 AUROC on clean engine audio, while phone-recorded audio remains a harder setting.
- •The repository includes a pre-trained model, a synthetic demo clip, CLI commands, a web app, and documentation covering model metrics, architecture, and project limitations.