Fuzzy precision and recall measures for audio signals segmentation

The approach presented in this paper applies fuzzy set theory to the evaluation of audio signals segmentation with high resolution and accuracy. The method is based on comparing automatically found boundaries with ground truth. Hence, the method is more accurate and able to grasp the evaluation problem in a way more similar to the evaluation conducted by a human being. Traditional methods often fail on grading segmentation algorithms, particularly those of relatively similar qualities. We define a fuzzy membership function that measures the degree to which the segments obtained by an automatic procedure are similar to the results of a correct segmentation. To identify a pair of equivalent segments, we set a fuzzy alignment function that points the pairs of segments obtained by an automatic segmentation with the corresponding segments from a correct segmentation. Speech segmentation is an example where the presented approach was applied.