Building a state-of-the-art audio classifier through machine learning

Sruthi Kurada, Massachusetts, USA 13-15

The availability of highly accurate audio classifiers is necessary for their use in the real world. This study explored the development of reliable audio classifiers to classify environmental sounds from the UrbanSound8k dataset. Published classifiers on this dataset only have 50%–74% accuracy range. I used some of the popular machine learning techniques such as feature selection, data preprocessing, data augmentation, and classifier parameter tuning to build a classifier with 99.4% accuracy. The steps used in the development of this state-of-the-art classifier are readily applicable to create highly accurate classifying machines in other audio-problem domains. As a proof, by applying the same techniques to a different dataset, normal vs. abnormal heart sound recordings, I have developed another high-performing classifier with 98% accuracy.

Show more