Big Chord Data Extraction and Mining

摘要: Harmonic progression is one of the cornerstones tonal music composition and thereby essential to many musical styles traditions. Previous studies have shown that genres composers could be discriminated based on chord progressions modeled as n-grams. These were however conducted small-scale datasets using symbolic transcriptions. In this work, we apply pattern mining techniques over 200,000 sequences out 1,000,000 extracted from I Like Music (ILM) commercial audio collection. The ILM collection spans 37 includes pieces released between 1907 2013. We developed a single program multiple data parallel computing approach whereby feature extraction tasks are split up run simultaneously cores. An audio-based recognition model (Vamp plugin Chordino) was used extract set. To keep low-weight sets, stored compact binary format. CM-SPADE algorithm, which performs vertical sequential patterns co-occurence information, fast efficient enough applied big collections like In orderto derive key-independent frequent patterns, transition chords by changes qualities (e.g. major, minor, etc.) root keys fourth, fifth, etc.). resulting vary in length (from 2 16) frequency 19,820) across genres. As illustrated graphs generated represent 4-chord progressions, some circle-of-fifths movements well represented most but varying degrees. These large-scale results offer opportunity uncover similarities discrepancies sets therefore build classifiers for search recommendation. They also support empirical testing theory. It more difficult new hypotheses such dataset due its size. This can addressed detection algorithms or suitable visualisation present companion study.