Evaluation of Musical Feature Extraction Tools using Perceptual Ratings.
Anton Hedblad · 2011
The increasing availability of digital music has created a demand for organizing and retrieving the music. Thus, a new multi-disciplinary research area called music information retrieval, MIR, has emerged. An important part of the content-based field of the research area is to extract musical features, such as tempo or modality, directly from the content, i.e. the audio.This thesis is an evaluation of available musical feature extraction tools. The evaluation is done by using the extracted musical features as predictors for perceptual ratings in correlation and regression analyses. The 11 perceptual ratings were gathered from a listening test. 22 musical audio features were extracted from the same stimuli, using different systems for musical feature extraction. These were chosen to predict some of the perceptual ratings based on findings in literature.High inter-subject reliability in the listening test implied a high agreement among the subjects, indicating that 20 subjects were enough. Fairly low inter-correlations between the ratings indicated that they were rated independently from each other. Six out of seven perceptual ratings with a priori selected predictors correlated moderately with their corresponding predictor (r>0.6). The results from the stepwise regression analyses were also moderate, where the amount of variance explained by the predictors ranged from 29-76%, indicating that there is room for improvements for developing new feature extraction algorithms.