Evaluation of similarity measures for spectro-temporal grouping with receiver operating characteristics
* Presenting author
In order to segregate different sources from a mixture of sounds, it is useful to decompose the mixture into smaller time-frequency units and estimate which source is dominating within the regarded segment. To determine the dominant source, acoustic features such as periodicity or binaural cues are extracted for each of the small units. By comparing the features of neighboring time-frequency units, it can be estimated whether the two should be considered as being dominated by the same source (“joint”), or whether they should be treated as "disjoint" since they show rather different attributes. Thus, the estimation can be regarded as classification problem in which each transition between adjacent segments is classified as either joint or disjoint transition. As a classification criterion, the similarity between specific acoustic features (which reflect the characteristics of the dominant source) can be used. Such a similarity measure can e.g. be the correlation coefficient.This study is concerned with the suitability of different similarity measures for the classification of next-neighbor transitions based on periodicity and binaural features. By means of receiver operating characteristics, several measures are evaluated to determine how useful they are for source segregation.