Examining Musical Meaning in Similarity Thresholds

Many approaches to Music Information Retrieval tasks rely on correctly determining if two segments of a given musical recording are repeats of each other. Repetitions in recordings are rarely exact, and identifying the appropriate threshold for these pairwise decisions is crucial for tuning MIR algorithms. However, current approaches for determining and reporting this threshold parameter are devoid of contextual meaning and interpretations, which makes comparing previous results difficult and which requires access to specific datasets. This paper highlights weaknesses in current approaches to choosing similarity thresholds, provides a framework using the proportion of orthogonal musical change to tie thresholds back to feature spaces with the cosine dissimilarity measure, and introduces new research possibilities given a music-centered approach for selecting similarity thresholds.