To quantify differences of opinion based on attributes, we extracted from each shot four acoustic characteristics designed to capture aspects related to speed, rhythm, harmony and timbre. Our hypothesis was that if the hierarchical comments obtained small L-measures and the annotators were actually measured by different acoustic properties, this effect should be evident when comparing notes in a representation derived from acoustic characteristics. All audio files were mixed with 22,050 Mono Hz before the extraction of the features and all analyses were performed with librosa 0.5 dev (McFee et al., 2015b). To view the features described in this section, see Figure66. For the rest of this article, we summarize the agreement between two notes according to measurement F, using accuracy and recall for pair classification and over-segmentation and sub-segmentation for NCE-Metrics. The first comparison in the matrix is therefore accessibility and maintenance. Discuss within the group and come to a conclusion for this issue, which is more important among the two attributes. At this point, this couple is the only one to be compared. After comparing a couple, continue with the next couple. If both criteria are of the same importance, place both letters in the corresponding cell. Note that the comparisons are in pairs - we completely ignore the other criteria.
The results of Experiment 2 show that different types of listening were based on different acoustic characteristics of the music. Comparing differences in characteristic correlations can help identify potential causal factors that contribute to listeners` interpretation of the musical form. Functionality analysis provides objective evidence of qualitative observations on how and why listeners interpret the musical structure differently, particularly in cases of significant disagreements. The Structural Poly Annotations of Music is a collection of hierarchical annotations for 50 pieces of music that were recorded by five experts with notes. The notes contain coarse and fine levels of segmentation that follow the same guidelines that are used in salami. The music in the spam collection contains examples of the same styles as SALAMI. The tracks were automatically scanned from a larger collection based on the degree of segment agreement between a series of estimates made by different algorithms (Nieto and Bello, 2016). 45 of these tracks are particularly demanding for current automatic segmentation algorithms, while the other five are simpler in terms of border detection. In the current work, we treat all tracks in the same way and use the 10 comparison pairs between different annotators per track. The spam list contains the same audio examples as the SALAMI collection described above, but the annotators are different, so the notes are shared between the two collections. The L-shaped measurement is calculated by identifying three periods of time (t, u, v) where (t, u) meet at a deeper level of the hierarchy (indicated by continuous lines) such as (t, v) (dotted lines) as shown in the left diagram (Annotator 1). In this example, the left note M (t, u) - 2 (both belong to lower level segments under the name d) and M (t, v) - 1 (both belong to upper level segments labeled C).
The right note has M (t, u) - M (t, v) - 2: All three moments belong to the legend of the f-segment, as indicated by the continuous lines.
