Machine Sound Monitor: Hear Faults Before Failure

Machines rarely fail in silence. Weeks before a fan or motor breaks, it starts to sound different, but nobody is listening. Maintenance teams cannot stand next to every machine, so the change is usually noticed only when it is loud, and by then it means downtime and an expensive repair. Machine Sound Monitor turns a device that is already on site into a listener. Today that is any phone or laptop browser; an IoT device with a sound sensor can send audio to the same API with its own device key. No special hardware. Placed next to the machine, it spends about a minute learning that machine's own normal sound, then locks it in until someone asks it to re-learn. The device sends a 3-second window of sound every second. The API resamples it to 16 kHz and compares its sound character to the learned normal, not its volume. A single odd sound, like a door slam, is ignored: only when 3 of the last 5 windows are unusual does it open one alert, then it stays quiet until the machine recovers. The workspace members get an email, and every open dashboard sees the scores and the alert live, where someone can acknowledge it. When the signal is too weak, it says "Unable to assess" instead of guessing. Raw audio is never stored. On a real phone recording of a desk fan, normal sound scored 0.23 to 0.49 and a rattling defect 0.85 to 2.36, against a learned alert line of about 0.9. Built with React, TypeScript and the Web Audio API in the browser, a FastAPI backend with live streaming, Supabase sign-in and Postgres, and Resend for email. For industrial machines we also trained PyTorch autoencoders on healthy DCASE 2020 fan audio (ROC AUC 0.65, close to the official baseline), with a training pipeline packaged for AMD Instinct MI300X GPUs on ROCm. An early warning that asks a person to check, not a diagnosis or certified safety system.