Applying Audio Problem Detection Algorithms to Sounds on Freesound Web Platform

With the domination of music broadcasting through the internet, detecting audio problems on sound files became more important. Essentia and other similar sound libraries enabled to detect audio problems automatically, which made the detection process faster, cheaper, and more accurate. These algorithms were mostly applied on music files which are usually recorded with adequate equipment in studio environment, and afterwards that are mixed and mastered. But there are many types of sounds other than music that can be recorded and uploaded to web. Freesound is a web platform that has currently the largest collection of sounds that are recorded in different environments, using different equipment, and tagged with suitable keywords. FSD50K is the dataset that includes 51,197 carefully annotated sounds gathered from Freesound Platform.
In this study, five of the mostly used automatic audio problem detection algorithms were selected. These are discontinuity, gap, click, hum, and clipping problems. Using Essentia, these algorithms were applied on FSD50K. Some bugs in the algorithms were detected and fixed, or some limitations were identified in their usage. For each audio problem, the sound classes that have the most problematic audio file percentages were detected.
Besides these already implemented algorithms, a detection algorithm for asymmetry, that is another property of sound, and which does not always have to be a problem, was implemented and results were analyzed.
An audio analyzer was implemented and added to Freesound Platform that allows users to filter the sound they are looking for based on whether the sound is problematic or not, according to the audio problem detection algorithms tested in the study.