| Symposium, 29 February 2012: Harmony and variation in music information retrieval |
|
Venue How to get there? Registration Automatic harmonic analysis and its applications Music theory has been essential in composing and performing music for centuries. Within Western tonal music, from the early Baroque on to modern-day jazz and pop music, the function of chords within a chord sequence can be explained by harmony theory. Although Western tonal harmony theory is a thoroughly studied area, formalising this theory is a hard problem. In this talk article we present HarmTrace, a functional model of Western tonal harmony, which builds on well-known theories of tonal harmony. Given a sequence of symbolic chord labels, HarmTrace automatically derives the harmonic relations between chords. In contrast to many other theories which remain purely theoretical, we present an implemented system and show how such a system can aid in solving typical MIR tasks, such as similarity estimation and chord recognition from musical audio. Functional modelling of musical harmony Music scholars have been intensively studying tonal harmony for centuries, yielding numerous theories and models. Unfortunately, a large number of these theories are formulated in a rather informal fashion and lack mathematical precision. We present a formalisation of the rules of tonal harmony as a Haskell (generalized) algebraic datatype. Given a sequence of chord labels, the harmonic function of a chord in its tonal context is automatically derived. For this, we use several advanced functional programming techniques, such as type-level computations, datatype-generic programming, and error-correcting parsers. For illustrative purposes, we compare our Haskell approach to an earlier solution in Java. We conclude that using Haskell lead to a more efficient, expressive, and elegant solution, that is easy to explain, modify, and maintain.
Syntactic structure in music and its implications Although there are many textbooks on tonal harmony, there are relatively few descriptions of how harmonic sequences are built together in tonal music. In my contribution I will present reasons to assume that harmony is governed by context-free dependency relationships and discuss what it means to hear musical syntax, to hear trees and how these affect perceptual tasks. I will compare these theoretical points with evidence from cognitive experiments and computational models.
Automatically extracting harmony from recorded music — and what to do with it Automatically extracting harmony from recorded music is a difficult engineering task. I will give an brief overview of my own work in chord transcription, highlighting how modelling music theoretic concepts can improve the performance of harmonic analysis systems. However, I will focus on applications of automatic harmony analysis: SongPrompter (guitar karaoke), a YouTube chord analyser based on my code, and an experimental Spotify App called "Last.fm Driver's Seat".
A cross-version approach for harmonic analysis of music recordings The automated extraction of chord labels from audio recordings is a central task in music information retrieval. Here, the chord labeling is typically performed on a specific audio version of a piece of music, produced under certain recording conditions, played on specific instruments and characterized by individual styles of the musicians. As a consequence, the obtained chord labeling results are strongly influenced by version-dependent characteristics. In this presentation, we show that analyzing the harmonic properties of several audio versions synchronously stabilizes the chord labeling result in the sense that inconsistencies indicate version-dependent characteristics, whereas consistencies across several versions indicate harmonically stable passages in the piece of music. In particular, we show that consistently labeled passages indeed correspond to correctly labeled passages. Our experiments show that the cross-version labeling procedure significantly increases the precision of the result while keeping the recall at a relatively high level. Furthermore, we introduce a powerful visualization which reveals the harmonically stable passages on a musical time axis specified in bars. Finally, we demonstrate how this visualization facilitates a better understanding of classification errors and may be used by musicologists as an inspiring tool for exploring harmonic structures.
Genre classification in music: the role of intervals Although automatic classification of music has been a well-studied topic in the literature, it remains a difficult problem. Many related tools, for example for music recommendation (Apple: genius) have been introduced, however most of them are based on meta-data and are not looking into the music itself.
MUSIVA: Linking variation and similarity in music In this talk I discuss the musicological concept of variation in music as a base for defining similarity in music, as it is researched in the current MUSIVA-project at Utrecht University (“Modelling MUsical SImilarity over time through the VAriation principle”). Musical similarity is a central concept in MIR for searching large collections of digitized music. Yet, MIR-researchers point to the huge challenge that modelling musical similarity poses to them, referring to similarity as an “ill-posed problem”, a “fuzzy” term or an “elusive” concept. One of the biggest challenge for MIR is how musical features extracted from digitized musical documents can be related to the experience of similarity by human listeners. Studies in Musicology and Music Cognition suggest that the musicological concept of variation is linked to the experience of similarity in music both by experts and novices. Variation is considered as a universal principle in music, hence can deliver the base for a general concept of musical similarity. Addressing this goal, the MUSIVA-project at Utrecht University researchers computational methods to model variation in music in classical, folk and popular music.
Information technologies for the discovery of culture specific music repertoires Music information research relates to the development of technologies for data intensive processing in music applications. Current information technologies do not pay attention to culture specificity and most current research is not attempting to change that either. For the improvement of current music information technologies there is a clear need to take a cultural approach. We need to develop well annotated data repositories of specific musical cultures. We need to work on data processing approaches that focus on finding differences between musical repertoires. We have to find applications that are of relevance to specific cultures. This is the motivation behind a large and ambitious recently started project funded by the?European Research Council entitled "CompMusic: Computational Models for the discovery of the world's music." The project is studying five music cultures: Hindustani, Carnatic, Turkish-makam, Andalusi, and Han. In this presentation I will go over the ideas supporting this?project in the context of IT research, emphasizing the challenges that we want to work on, and the proposed approaches to tackle these challenges.
Motif-Based Recognition of Folk Song Melodies - a First Step One of the conclusions of the recent WITCHCRAFT-project is that recurring melodic motifs are of major importance for determining tune-family membership of folk-song melodies. The ethnomusicological concept of tune family denotes a group of melodies that presumably have a common historical origin. Through the process of oral transmission, considerable changes occur to the melodies over time, resulting in groups of variants. As part of the WITCHCRAFT-project, the musicological domain experts from the Meertens Institute annotated c. 1400 occurrences of motifs in a set of 360 melodies from 26 tune families. In the current Tunes & Tales project, we are employing these annotations to perform motif-based recognition of melodies. In this talk, I will introduce the main hypothesis of the Tunes & Tales project. Both folk-melodies and folk-tales will be formalized as layered sequences of motifs, which enables the modelling of oral transmission in terms of changes in these sequences. Next, I will present the first results of a melody classification experiment that is based on the annotated motifs. |