Relying on AI in music analysis: An evaluation of predictive consistency
Co-Authors:Cameron J. Anderson, Michael Schutz
Virtual or In-person:In-person
Music Information Retrieval (MIR) is a growing field focused on algorithmically extracting information from music. One popular application of MIR is developing automated musical analyses to analyze and estimate information from audio files. Extractions cover both concrete topics like genres, artists, and instrumentation, and abstract topics like tonality, patterns, and emotion (Diakopoulos et al., 2009; Li et al., 2017). MIRtoolbox, a prominent analysis library with over 1900 citations (Lartillot & Toiviainen, 2007; Lartillot et al. 2008), provides tools to aid in the extraction and manipulation of relevant musical features. Despite its extensive use, little research has investigated the consistency of MIRToolbox’s extraction algorithms across similar audio files. To address this issue, we analyzed consistency of tonality estimates across four performers’ interpretations of Chopin’s 24 Préludes, introducing a novel method of assessing the predictive consistency of key and mode. Key refers to the tonal center around which a piece of music is organized, influencing the frequency of notes. Key often implies a specific mode—a structural syntax of music, influencing the notes and sequences of a piece that contribute to the emotional aspect of music. Measuring consistency across different interpretations of the same pieces clarifies algorithmic behaviour when evaluating subjective aspects like mode prediction, where consistent note values should yield similar results. Results reveal inconsistent predictions in almost half the excerpts among different performers and perceptual inconsistencies of predictive mode. Following these limitations and capabilities of MIR algorithms, this study offers a pioneering approach to assessing evaluations of abstract information.