A Model-Based Approach to Constructing Music Similarity Functions

Abstract

Several authors have presented systems that estimate the audio similarity of two pieces of music through the calculation of a distance metric, such as the Euclidean distance, between spectral features calculated from the audio, related to the timbre or pitch of the signal. These features can be augmented with other, temporally or rhythmically based features such as zero-crossing rates, beat histograms, or fluctuation patterns to form a more well-rounded music similarity function. It is our contention that perceptual or cultural labels, such as the genre, style, or emotion of the music, are also very important features in the perception of music. These labels help to define complex regions of similarity within the available feature spaces. We demonstrate a machinelearning-based approach to the construction of a similarity metric, which uses this contextual information to project the calculated features into an intermediate space where a music similarity function that incorporates some of the cultural information may be calculated.

A Survey of Evaluation in Music Genre Recognition

“…We find that both Classify and Cluster are used in about 2.6% (12) of the experimental work [33,72,126,189,196,242,253,261,301,320,430,438]. A seventh experimental design is Retrieve, which appears in at least 19 works [10,46,57,61,86,118,119,121,203,222,232,262,320,348,384,388,446,447,466]. For instance, Kuo and Shan [203] [401] investigate the variability of their system using Classify by using features computed from excerpts of several durations.…”

“…The FoM most often reported in the case of the Retrieve experimental design is Precision@k. This FoM is reported in 12 of the 19 works using Retrieve [10,57,61,86,203,222,262,320,384,446,447,466]; and [388] reports "normal-ized precision" and "normalized recall," which takes into account the ranking of retrieved elements. Of the references using Retrieval, the ROC is reported in [121,466].…”

Singer verification: Singer model .vs. song model

“…If song-models are simple Gaussians, KL can be applied directly. If songs are modeled as GMM, the KL divergence can be approximated by the Monte Carlo method ( [1], [2]) or by the Earth Moving Distance ( [11], [13], [14]). The difference between the two approaches is not significant in term of quality according to [10].…”

Unifying Low-Level and High-Level Music Similarity Measures

“…The evaluation includes the embeddings, which merge timbral and tonal distances, and, alternatively, timbral and semantic distances. West and Lamere [10] apply classifiers to infer semantic features of the songs. In their experiment, Mel-frequency spectral irregularities are used as an input for a genre classifier.…”

“…The second idea we explore shifts the problem to a more high-level (semantic) domain as we propose to use high-level semantic dimensions, including information about genre and musical culture, moods and instruments, and rhythm and tempo. With regard to this aspect, we continue the research of [8]- [10] but, more in the line of [10], we investigate the possibility of benefiting from results obtained in different classification tasks and transferring this acquired knowledge to the context of music similarity (Sec. III-B).…”