Master Assignment
Aggression Based Ranking of Audio Samples; Annotation and Automatic Ranking
Type: Master M-ITECH
Location: University of Twente
Duration: Dec, 2018 - Aug, 2019
Student: Wiltenburg, (Daan, Student M-ITECH)
Date Final project: August 15, 2019
Supervisors:
Abstract:
In many places, public and private, aggression can be a big problem. Therefore, surveillance is essential. This paper describes the research in automatic audio ranking based on aggression. The research is two-fold: First the most efficient way of rank annotation is determined by comparing different methods. A combination of pair-wise comparison and binary insertion sort turns out to be the most time efficient approach, while at the same time yielding the strongest ranking. The second problem addressed in this research is in the field of machine learning and learning-to-rank. Two type of loss functions are compared. The first loss function is Mean Squared Error, which uses a continuous target between 0 and 1 to train a network making this a regression approach. The second loss function is the log-likelihood function, which takes an ordered list as target making this a list-wise approach. The regression method performs comparable to human annotators, with a Kendall’s Rank Correlation Coefficient of 0.8228, where the human annotators achieved a Kendall’s Rank Correlation Coefficient of 0.88.