My thesis at Gatech was ‘Utilizing I-Vectors to classify fault sounds in machinery using ToyADMOS data set’ under supervision of Dr. David Anderson.
ToyADMOS was a recently released data set by Koizumi. et all which can be found at link. ADMOS stands for anomaly detection in machine operating sounds. It is much harder to get data for sounds of faulty machinery because ideally machines are made to be robust. And it can take a long time to collect enough data of anomalous scenarios to use in machine learning.
Traditionally an AE or similar mechanism is used to model the normal data, and anything deviating too much from this is classified as anomalous. However, with anomalous data present, it opens the possibility to train models to reduce False Positives while increasing True Positives simultaneously, not necessarily an inherent quality of traditional methods which may end up being too generalized.
We used I-Vectors (link for further details). I-vectors are being used primarily for speech detection, but we were able to achieve positive results for our use case.