Audio Problems Detection on Sound Collections

TitleAudio Problems Detection on Sound Collections
Publication TypeMaster Thesis
Year of Publication2019
AuthorsBadenas, V.
AbstractThis thesis is a first approach to creating and evaluating algorithms whose aim is to detect an unwanted feature in audio streams. Existing algorithms are evaluated and it’s parameters optimized for audio files without a specific length. These algorithms are used to detects defects such as clipping/saturation, Clicks, Noise Bursts, Hum, etc. Additionally, algorithms are developed in this thesis and also evaluated. To evaluate those results, a subset of Freesound is used after being annotated manually using scripts developed by the author. This ground truth is further used to evaluate the performance of the algorithms against it and is also used to sweep different values of the parameters available in the algorithms to find the parametrization that gives the best performance over a dataset of audios with multiple problems. For the evaluation dataset, the train set for the Kaggle Audio Tagging 2019 competition is used, as it is a collection of sounds with many problems to be detected. Due to some issues and time constraints, the results are not as satisfactory as previously expected.
KeywordsAudio defects Detection, Audio Problems, Audio tagging
intranet