̽����ѡ

This paper presents a multimodal dataset of Greek folk dance music, focusing on syrtos and balos. Developed to support research in computational musicology, the dataset improves access to Greek musical heritage through manually transcribed MIDI scores, aligned lyrics, and rich metadata, all curated by expert musicologists. Through pattern analysis and feature extraction, we examine both shared melodic structures and unique characteristics of each dance, with some examples reflecting traces of oral transmission. While metadata accompanies the collection to support organization and context, our primary emphasis is on the musical and lyrical content. This work contributes to digital ethnomusicology by showing how multimodal datasets of folk music can inform both analytical research and cultural heritage preservation.

Music question–answering (MQA) is a machine learning task where a computational system analyzes and answers questions about music‑related data. Traditional methods prioritize audio, overlooking visual and embodied aspects crucial to music performance understanding. We introduce MusiQAl, a multimodal dataset of 310 music performance videos and 11,793 human‑annotated question–answer pairs, spanning diverse musical traditions and styles. Grounded in musicology and music psychology, MusiQAl emphasizes multimodal reasoning, causal inference, and cross‑cultural understanding of performer–music interaction. We benchmark AVST and LAVISH architectures on MusiQAI, revealing strengths and limitations, underscoring the importance of integrating multimodal learning and domain expertise to advance MQA and music information retrieval.

Visualizing music—through music notation, analytical representations, or music videos—might potentially boost the appreciation of music in all its richness. The purpose of this study was to design and test a visualization strategy aimed at explicating to a large audience with diverse backgrounds—especially novices—the multifaceted beauty of the final Contrapunctus in J.S. Bach's The Art of Fugue, performed by the Danish String Quartet. At the surface level of the musical structure, the rich fluctuation of pitch shaped by each musician was depicted in the form of undulating pitch curves. At a deeper structural level, the repetition of pitch curves, distinctive of fugues, was highlighted through vertical alignment—inspired by a technique called paradigmatic analysis, originating from anthropology and music semiology. The visualization was initially prototyped in the form of a real-time technology as part of the MusicLab Copenhagen research concert. The concert audience focused on the performance itself, and did not pay much attention to, nor appreciate, the visualization. To evaluate more thoroughly the potential of the visualization, participants with varied musical expertise and taste were invited to listen to a recorded performance of the piece and watch the visualization on their own computer. A large majority reported that they felt they understood the visualization, around half of them felt that it enhanced their musical understanding, and a small group felt that it helped them to better appreciate the music.

MusicLab Copenhagen was a unique research concert featuring the world-renowned Danish String Quartet in a naturalistic setting. The audience was split between one group physically located in the hall, another group listening to a radio broadcast, and a third group watching a live stream. Qualitative and quantitative data were captured from both musicians and audiences, resulting in a comprehensive dataset that can be used to address many research questions. This document introduces the dataset, explains its structure, and reflects on the related data collection, storing, publishing, and archiving processes.

This paper addresses improving performance analysis by automating the recognition of expressive performance styles. We propose a multimodal fusion approach integrating audio, video, and motion data. We demonstrate the effectiveness of our approach by utilizing convolutional neural network (CNN) models. Training is done on a classical piano dataset of 211 excerpts containing audio, video, MIDI, and motion capture data. The results highlight the robustness of the CNN models; they achieve high accuracy even when trained on a limited dataset. Our study contributes to advancing the field of performance analysis by applying deep learning techniques to multimodal data.

The TIME project: Timing and Sound in Musical Microrhythm (2017–2022) studied microrhythm; that is, how dynamic envelope, timbre, and center frequency, as well as the microtiming of a variety of sounds, affect their perceived rhythmic properties. The project involved theoretical work regarding the basic aspects of microrhythm; experimental studies of microrhythm perception, exploring both stimulus features and the participants’ enculturated expertise; observational studies of how musicians produce particular microrhythms; and ethnographic studies of musicians’ descriptions of microrhythm. Collectively, we show that: (a) altering the microstructure of a sound (“what” the sound is) changes its perceived temporal location (“when” it occurs), (b) there are systematic effects of core acoustic factors (duration, attack) on microrhythmic perception, (c) microrhythmic features in longer and more complex sounds can give rise to different perceptions of the same sound, and (d) musicians are highly aware of microrhythms and have developed vocabularies for describing them. In addition, our results shed light on conflicting results regarding the effect of microtiming on the “grooviness” of a rhythm. Our use of multiple, interdisciplinary methodologies enabled us to uncover the complexity of microrhythm perception and production in both laboratory and real-world musical contexts.

Analysing electroacoustic music remains challenging, leaving this artistic treasure somewhat out of reach of mainstream musicology and many music lovers. This chapter examines electroacoustic music analysis, covering musicological investigations and desires and technological challenges and potentials. The aim is to develop new technologies to overcome the current limitations. The compositional and musicological foundations of electroacoustic music analysis are based on Pierre Schaeffer’s Traité des objects musicaux. The chapter presents an overview of core analytical principles underpinning more recent musicological approaches, including R. Murray Schafer’s soundscape analysis, Denis Smalley’s spectro-morphology, and Lasse Thoresen’s graphical formalisation. Then the state of the art in computational analysis of electroacoustic music is compiled and organised along broad themes, from detecting sound objects to estimating dynamics, facture and grain, mass, motions, space, timbre and rhythm. Finally, I sketch the principles of what could be a Toolbox des objets sonores.

The term “multimodal music dataset” is often used to describe music-related datasets that represent music as a multimedia art form and multimodal experience. However, the term “multimodality” is often used differently in disciplines such as musicology, music psychology, and music technology. This paper proposes a definition of multimodality that works across different music disciplines. Many challenges are related to constructing, evaluating, and using multimodal music datasets. We provide a task-based categorization of multimodal datasets and suggest guidelines for their development. Diverse data pre-processing methods are illuminated, highlighting their contributions to transparent and reproducible music analysis. Additionally, evaluation metrics, methods, and benchmarks tailored for multimodal music processing tasks are scrutinized, empowering researchers to make informed decisions and facilitating cross-study comparisons.

Norske folkemusikkarkiver står foran utfordringer vedrørende å gjøre sine samlinger tilgjengelige på nett når feltopptak faller i det fri etter 50 år. Denne artikkelen beskriver et forsøk på å tilrettelegge brukskopier i form av lydfiler slik at brukerne kan lytte til enkeltopptak i en nettpresentasjon av et arkivs innhold. Lydklassifiseringsteknologi er i stand til å finne og markere starttidspunkt på enkeltmelodier og spare arkivpersonalet for mange timers manuelt arbeid. Mirage-prosjektet på UiOs RITMO-senter har utviklet et grensesnitt for et slikt verktøy for Nasjonalbibliotekets folkemusikkarkiv og dets nettkatalog WebbFIOL. Løsningen vil kunne tas i bruk av alle som står overfor liknende utfordringer.

We present an ongoing project dedicated to the transmutation of a collection of field recordings of Norwegian folk music established in the 1960s into an easily accessible online catalogue augmented with advanced music technology and computer musicology tools. We focus in particular on a major highlight of this collection: Hardanger fiddle music. The studied corpus was available as a series of 600 tape recordings, each tape containing up to 2 hours of recordings, associated with metadata indicating approximate positions of pieces of music. We first need to retrieve the individual recording associated with each tune, through the combination of an automated pre-segmentation based on sound classification and audio analysis, and a subsequent manual verification and fine-tuning of the temporal positions, using a home-made user interface. Note detection is carried out by a deep learning method. To adapt the model to Hardanger fiddle music, musicians were asked to record themselves and annotate all played note, using a dedicated interface. Data augmentation techniques have been designed to accelerate the process, in particular using alignment of varied performances of same tunes. The transcription also requires the reconstruction of the metrical structure, which is particularly challenging in this style of music. We have also collected ground-truth data, and are conceiving a computational model. The next step consists in carrying out detailed music analysis of the transcriptions, in order to reveal in particular intertextuality within the corpus. A last direction of research is aimed at designing tools to visualise each tune and the whole catalogue, both for musicologists and general public.

This article addresses the computational estimation of attack regions in audio recordings. Previous attempts to do so were based on the reduction of the audio waveform into an envelope curve, which decreases its temporal resolution. The proposed approach detects the attack region directly from the audio waveform. The attack region is modeled as a line starting from a low-amplitude point and intersecting one of the local maxima according to two principles: (1) maximizing the slope, while favoring, at the same time, a higher peak if the slope remains only slightly lower and (2) dismissing initial attack regions of relatively low amplitude. The attack start position is fine-tuned by intersecting the attack slope with the audio waveform. The proposed method precisely pinpoints the attack region in cases where it is unambiguously observable from the waveform itself. In such cases, previous methods selected a broader attack region due to the loss of temporal resolution. When attack regions are less evident, the proposed method’s estimation remains within the range of results provided by other methods. Applied to the prediction of judgments of P-center localization [Danielsen, Nymoen, Anderson, C^amara, Langerød, Thompson, and London, J. Exp. Psychol. Hum. Percept. Perform. 45, 402–418 (2019)], the proposed method shows a significant increase in precision, at the expense of recall.

I present a short overview of computational methods for musicological analysis of notated music. We first need to clarify the various levels of computational representations of music: on one side, notated music, on the other, audio recordings, and in the middle, a note-level representa- tion of music performance where higher-level musical descriptions are absent. The article provides a synthetic and partial panorama of the different types of music analysis that have been systematised and auto- mated using computers. While pioneering works were mainly focused on statistical descriptions of the surface of music, other dimensions of music analysis such as harmony, metre and structure have been taken into consideration since. I conclude by sketching my personal vision of the future of computational music analysis.

This paper presents a Hardanger fiddle dataset “HF1” with polyphonic performances spanning five different emotional expressions: normal, angry, sad, happy, and tender. The performances thus cover the four quadrants of the activity/valence-space. The onsets and offsets, together with an associated pitch, were human-annotated for each note in each performance by the fiddle players themselves. First, they annotated the normal version. These annotations were then transferred to the expressive performances using music alignment and finally human-verified. Two separate music alignment methods based on image registration were developed for this purpose; a B-spline implementation that produces a continuous temporal transformation curve and a Demons algorithm that produces displacement matrices for time and pitch that also account for local timing variations across the pitch range. Both methods start from an “Onsetgram” of onset salience across pitch and time and perform the alignment task accurately. Various settings of the Demons algorithm were further evaluated in an ablation study. The final dataset is around 43 minutes long and consists of 19 734 notes of Hardanger fiddle music, recorded in stereo. The dataset and source code are available online. The dataset will be used in MIR research for tasks involving polyphonic transcription, score alignment, beat tracking, downbeat tracking, tempo estimation, and classification of emotional expressions.

é

This paper develops the hypothesis that symbolic drum patterns can be represented in a reduced form as a sim- ple oscillation between two states, a Low state (commonly associated with kick drum events) and a High state (often associated with either snare drum or high hat). Both an onset time and an accent time is associated to each state. The systematic inference of the reduced form is formal- ized. This enables the specification of a rhythmic struc- tural similarity measure on drum patterns, where reduced patterns are compared through alignment. The two-state representation allows a low computational cost alignment, once the complex topological formalization is fully taken into account. A comparison with the Hamming distance, as well as similarity ratings collected from listeners on a drum loop dataset, indicates that the bistate reduction enables to convey subtle aspects that goes beyond surface-level com- parison of rhythmic textures.

The GrooveToolbox is a new Python toolbox implementing various algorithms, new and pre-existing, for the analysis and comparison of symbolic drum loops, including rhythm features, similarity metrics and microtiming features. As part of the GrooveToolbox we introduce two new metrics of rhythm similarity and four features for describing the significant properties of microtiming deviations in drum loops. Based on a two-part perceptual evaluation, we show these four new microtiming features can each correlate to similarity perception, and be used with rhythm similarity metrics to improve personalized similarity models for drum loops. A new measure of structural rhythmic similarity is also shown to correlate more strongly to similarity perception of drum loops than the more com- monly used Hamming distance. These results point to the potential application of the GrooveToolbox and its new features in drum loop analysis for intelligent music production tools. The GrooveToolbox may be found at: https://github.com/fredbru/GrooveToolbox

This paper reports on two experiments that investigated the expressive means through which musicians well versed in groove-based music signal the intended timing of a rhythmic event. Data were collected from 21 expert electric guitarists and 21 bassists, who were instructed to perform a simple rhythmic pattern in three different timing styles—“laidback,” “on-the-beat,” and “pushed”—in tandem with a metronome. As expected, onset and peak timing locations corresponded to the instructed timing styles for both instruments. Regarding sound, results for guitarists revealed systematic differences across participants in the duration and brightness [spectral centroid (SC)] of the guitar strokes played using these different timing styles. In general, laid-back strokes were played with a longer duration and a lower SC relative to on-the-beat and pushed strokes. Results for the bassists indicated systematic differences in intensity (sound-pressure level): pushed strokes were played with higher intensity than on-the-beat and laid-back strokes. These results lend further credence to the hypothesis that both temporal and sound-related features are important indications of the intended timing of a rhythmic event, and together these features offer deeper insight into the ways in which musicians communicate at the microrhythmic level in groove-based music.

We present a new system for real-time visualisation of music performance, focused for the moment on a fugue played by a string quartet. The basic principle is to offer a visual guide to better understand music using strategies that should be as engaging, accessible and effective as possible. The pitch curves related to the separate voices are drawn on a space whose temporal axis is normalised with respect to metrical positions, and aligned vertically with respect to their thematic and motivic classification. Aspects related to tonality are represented as well. We describe the underlying technologies we have developed and the technical setting. In particular, the rhythmical and structural representation of the piece relies on real-time polyphonic audio-to-score alignment using online dynamic time warping. The visualisation will be presented at a concert of the Danish String Quartet, performing the last piece of The Art of Fugue by Johann Sebastian Bach.

We present a method for tempo estimation from audio recordings based on signal processing and peak tracking, and not depending on training on ground-truth data. First, an accentuation curve, emphasizing the temporal location and accentuation of notes, is based on a detection of bursts of energy localized in time and frequency. This enables the detection of notes in dense polyphonic texture, while ignoring spectral fluctuation produced by vibrato and tremolo. Periodicities in the accentuation curve are detected using an improved version of autocorrelation function. Hierarchical metrical structures, composed of a large set of periodicities in pairwise harmonic relationships, are tracked over time. In this way, the metrical structure can be tracked even if the rhythmical emphasis switches from one metrical level to another. This approach, compared to all the other participants to the Music Information Retrieval Evaluation eXchange (MIREX) Audio Tempo Extraction competition from 2006 to 2018, is the third best one among those that can track tempo variations. While the two best methods are based on machine learning, our method suggests a way to track tempo founded on signal processing and heuristics-based peak tracking. Moreover, the approach offers for the first time a detailed representation of the dynamic evolution of the metrical structure. The method is integrated into MIRtoolbox, a Matlab toolbox freely available.

This paper describes a multidisciplinary encounter with oral testimony archives and their incorporation in the artistic research project Intimal. It explores ways in which to creatively listen to stories which might be emotionally challenging, and to create and sustain empathy with these disembodied voices, which contain a shared history as fissures of sixty years of violence.

Improvisation in music therapy has been shown to be an effective technique for engaging clients in emotionally rooted (inter)action to treat affective disorders such as major depression (Aalbers et al., 2017; Erkkilä et al., 2011). During improvisation, however, a variety of musical information is exchanged, resulting in a highly complex musical and interpersonal situation. While traditional models of music therapy analysis emphasise aural analysis and assessment of single sessions (Bruscia, 1987), more recent and elaborated methods, such as microanalysis, focus on the detailed development of improvisation sessions (Wosch, 2021; Wosch & Erkkilä, 2016), which comes at the cost of a more time-consuming application process. Digital processing, as in music information retrieval and machine learning, seems promising to accelerate the analysis process, but requires considerable preliminary work in data preprocessing and formalisation of the high-level concepts used in music therapy to develop a suitable dataset for model training. Moreover, additional benefits of digital processing comprehend a more detailed and precise analysis of musical data.

Music analysis involves categorising and interpreting sonic elements to uncover the structure and meaning of a work. In contemporary music studies, analysts often face methodological challenges in this process, especially when dealing with works that contain high degrees of complexity and ambiguity in terms of timbre, texture and temporal structure. This paper proposes a methodological model for analysing spatiotemporal complexities commonly observed in contemporary repertoires, utilising computational tools to enhance auditory memory and expand interpretative possibilities. Auditory memory plays a pivotal role in aural analysis, an approach that serves as a valuable alternative or complement to traditional score-based analysis. Rooted in Pierre Schaeffer’s typomorphology of objets sonores and the work of other analysts in electroacoustic music studies, the general principles of aural analysis can be outlined in a three-step process: 1) attentive listening to the acoustic properties of sounds, 2) describing and categorising their sonic variations, and 3) assessing their functions within a large-scale formal structure. Computational sound visualisation tools are frequently employed in this process to assist in transcribing and retaining musical events that are either absent from the score or difficult to interpret aurally due to textural complexities and/or timbral elusiveness. Despite their increasing use, however, the full potential of these tools remains largely unexplored in contemporary music studies. By digitally decomposing the transformation processes of ambiguous musical flows and supporting the organisation and structuring of auditory memory, computational analysis of audio data and various visualisation methods can deepen our understanding of both local sonic morphology and large-scale formal trajectory. In line with these considerations, the paper investigates how specialised computer interfaces can facilitate music analytical processes. Two research questions guide this investigation: 1) How can we analyse a stream of sonic textures; and 2) How can we outline the formal structure of a work that embraces extremes of sonic energy and polyrhythmic intricacy? To explore these questions, we have developed muScope, a new computer program that enables users to browse within high-resolution sonograms in tandem with a range of graphical representations capturing audio, timbral, rhythmic and structural descriptions. The analysis of spectral “fluctuations” allows for the identification of rapid pulsations at the middle ground between rhythm and timbre. Self-similarity matrix representations can serve as a tool for outlining the structural division of the audio data based on various sonic attributes. We integrate these visual representations into an analytical workflow designed to support the construction of a composition’s formal structure. Our methods are demonstrated through an analysis of excerpts from Kaija Saariaho’s Io for large ensemble and electronics (1986–87) and Raphaël Cendo’s Corps for piano and ensemble (2015). This integrated analytical approach offers new insights into the interplay between musical perception, memory and analytical interpretation using digital tools.

Background Decades of research in computational sound and music analysis has led to a large range of analysis tools offering rich and diverse description of music, although a large part of the subtlety of music remains out of reach. These descriptors are used to establish computational models predicting perceived or induced emotion directly from music. Although the models can predict a significant amount of variability of emotions experimentally measured (Panda et al., 2023), further progress seems hard to achieve, probably due to the subtlety of music and of the mechanisms underlying the evocation of emotion from music. Aims An extensive but synthetic panorama of computational research in sound and music analysis as well as emotion prediction from music is presented. Core challenges are highlighted and prospective ways forward are suggested. Main contribution For each separate music dimension (dynamics, timbre, rhythm, tonality and mode, motifs, phrasing, structure and form), a synthetic panorama of the state of the art is evoked, highlighting strengths and challenges as well as indicating how particular sound and music features have been found to correlate with rated emotions. The various strategies for modelling emotional reactions to audio and musical features are presented and discussed. One common general analytical approach carries out a broad and approximate analysis of the audio recording based on simple mathematical models, describing individual audio or musical characteristics numerically. It is suggested that such loose approach might tend to drift away from commonly understood musical processes and to generate artefacts. This vindicates a more traditional musicological approach based on a focus on the score or approximations of it – through automated transcription if necessary – and a reconstruction of the types of traditional representations commonly studied in musicology. I also argue for the need to closely reflect the way humans listen to and understand music, inspired by a cognitive perspective. Guided by these insights, I sketch the idea of a complex system made of interdependent modules, founded on sequential pattern inference and activation scores not based on statistical sampling. I also suggest perspectives for the improvement of computational prediction of emotions evoked by music. Discussion and conclusion Further improvements of computational music analysis methods, as well as emotion prediction, seem to call for a change of modelling paradigm. References R. Panda, R. Malheiro, R. Paiva, "Audio Features for Music Emotion Recognition: A Survey", IEEE Transactions on Affective Computing, 14-1, 68-88, 2023.

Jeg presenterer en rekke verktøy utviklet i samarbeid med Nasjonalbiblioteket. AudioSegmentor deler automatisk båndopptak i individuelle musikkstykker. Dette verktøyet forenklet digitaliseringen av Norsk folkemusikksamling. Vi bruker avanserte dyp læringsmetoder for å skape et banebrytende automatisk musikktranskriberingssystem, MusScribe, først finjustert for Hardingfele, og nå gjort tilgjengelig for musikkarkivprofesjonelle for et bredt spekter av musikk. Jeg diskuterer også våre pågående fremskritt innen den automatiserte musikologiske analysen av folkemusikkstykker og omfattende samlinger.

One aim of the MIRAGE project is to conceive new technologies allowing to better access, understand and appreciate music, with a particular focus on Norwegian folk music. This seminar presents what has been achieved during the four years of the project, leading in particular to the digital version of the Norwegian Catalogue of Folk Music. We are also conceiving tools to automatically transcribe audio recordings of folk music. More advanced musicological applications are discussed as well. To conclude, we introduce the new spinoff project, called muScribe, aimed at the development of transcription services, for a broad range of music, besides folk music, in a first stage tailored to professional organisations such as archives, publishers and producers.

My work involves developing computational tools to safeguard and elevate the cultural significance of music repertoires, with a focus on a cooperative project with the National Library of Norway related to their collection of Norwegian folk music. Our first phase centered on transforming unstructured audio tapes into a systematic dataset of melodies while ensuring its access and longevity through efficient data management and linking with other catalogues. Our core activity involves transcribing audio recordings into scores, comparing the traditional manual method with our modern attempts towards automation. Providing detailed performance notation, the close alignment between scores and audio recordings will help improve comprehension and overall accessibility, as well as a more advanced structuring of the collection. Challenges arose when incorporating this music into the International Inventory of Musical Sources (RISM) database due to the incompatible 'incipit' concept, unfitting genres like Hardanger fiddle folk music. We suggest innovative generalisations for this concept. Moreover, we're creating techniques to digitally dissect the musical corpus, aiming to extract key features of each tune. This initiative not only serves as an alternative to incipits but also provides novel metadata formats, increasing the usability and connectivity within its content and with other databases.

Computational music analysis, still in its infancy, lacking overarching reliable tools, can be seen at the same time as a promising approach to fulfill core epistemo- logical needs. Analysis in the audio domain, although approaching music in its entirety, is doomed to superficiality if it does not fully embrace the underlying symbolic system, requiring a complete automated transcription and scaffolding of metrical, modal/harmonic, voicing and formal structures on top of the layers of elementary events (such as notes). Automated transcription enables to get over the polarity between sound and music notation, providing an interfacing semiotic system that combines the advantages of both domains, and surpassing the limitation of traditional approaches based on graphic representations. Deep learning and signal processing approaches for the discretisation of the continuous signal are compared and discussed. The multi-dimensional music transcription and analysis framework (where both tasks are actually deeply intertwined) requires to take into account the far-reaching interdependencies between dimensions, for instance between motivic and metrical analysis. We propose an attempt to build such a comprehensive framework, founded on general musical and cognitive principles and an attempt to build music analysis capabilities through a combina- tion of simple and general operators. The validity of the analyses is addressed in close discussion with music experts. The potential capability to produce valid analyses for a very large corpus of music would make such a complex system a potentially relevant blueprint for a cognitive modelling of music understanding. We try to address a large diversity of music cultures and their specific challenges: among others, maqam modes (with Mondher Ayari), Norwegian Hardanger fiddle rhythm (with Mats Johansson and Hans-Hinrich Thedens), djembe drumming from Mali (with Rainer Polak) or electroacoustic music (Towards a Toolbox des objets musicaux, with Rolf Inge Godøy). We aim at making the framework fully transparent, collaborative and open.

The 2nd MIRAGE Symposium covers a broad range of topics related to the MIRAGE project, mainly related to music and emotion, music cognition in general, music analysis and music therapy. Featuring two keynotes by Patrik Juslin and Didier Grandjean.

As a prélude for Norway's Constitution Day, this special event celebrated the Norwegian folk music tradition, showcasing our new online archive and demonstrating the richness of Hardanger fiddle music, with live performance. One aim of the project is to conceive new technologies allowing to better access, understand and appreciate Norwegian folk music. In this event, we introduced a new online version of the Norwegian Folk Music Archive and discuss underlying theoretical and technical challenges. A live concert/workshop, with the participation of Olav Luksengård Mjelva, offered a lively introduction to Hardanger fiddle music and its elaborate rhythm. The interests and challenges of automated transcription and analysis were discussed, with the public release of our new software Annotemus. The symposium was organised in the context of the MIRAGE project (RITMO, in collaboration with the National Library of Norway's Digital Humanities Laboratory).

Music is something that we mostly all appreciate, yet it remains a hidden and enigmatic concept for many of us. Music notation, in the form of music scores, facilitates practicing and enhances the understanding of the richness of musical works. However, acquiring musical scores for any music performance is a tedious and demanding task (called music transcription) that demands considerable proficiency. Hence the interest of computational automation. But music is not just notes, it is also melody, rhythm, themes, timbre, and very subtle aspects such as form. While many of us may not be consciously familiar with these concepts, they still have a subconscious influence on our aesthetic experience. Interestingly, it often happens that the more we consciously understand the underlying language of music, the more we tend to appreciate and enjoy it. Therefore, there is value in creating computational tools that can automate and enhance these types of analyses. The presenters' past work resulted in the creation of Matlab's MIRtoolbox, which measures a broad range of musical characteristics directly from audio through signal processing techniques. Currently, the MIRAGE project prioritises music transcription (with a particular focus on Norwegian folk music), blending neural-network-based deep learning with conventional rule-based models. Through this project, they highlight the importance of acknowledging the interconnectedness between all musical elements. Additionally, they have crafted animated visualisations to make analyses more accessible to the general public and are aiming to make music transcription technology available to the public, with support from UiO Growth House.

In the rapidly expanding field of music information retrieval (MIR), automatic transcription remains one of the most sought-after capabilities, especially for songs that employ multiple instruments. Musscribe emerges as a state-of-the-art transcription tool that addresses this challenge by integrating three distinct methodologies: demixing, harmonic dilated convolution, and joint beat tracking. Demixing is employed to isolate individual instruments within a song by separating overlapping audio sources, thus ensuring each instrument is transcribed distinctly. Beat tracking is then run as a parallel process to extract the joint beat and downbeat estimations. These processes results in an output midi file, which is then quantized using information derived from the beat tracking. As such, this method paves the way for more accurate and sophisticated analyses, bridging the gap between human and machine understanding of music. Together, these methodologies allow us to produce transcriptions that are not only accurate but also highly representative of the original compositions. Preliminary tests and evaluations showcase the potential in transcribing complex musical pieces with high fidelity, outperforming many contemporary tools in the market. This innovative approach not only has implications for music transcription but also for broader applications in audio analysis, remixing, and digital music production. The model has been instrumental in accelerating the composition process for several Norwegian television shows. Moreover, its efficacy can be observed in the Netflix series "A Storm for Christmas." Renowned composer Peter Baden harnessed this tool to enhance his workflow, proving the demand for innovative tools like this in the professional music industry.

Computational detection and analysis of sound objects is of high importance both for musicology and sound design. Yet Music Information Retrieval technologies have so far been mostly focusing on transcription of music into notes in a classical sense whereas we are interested in detecting sound objects and their feature categories, as was suggested by Pierre Schaeffer’s typology and morphology of sound objects in 1966, reflecting basic sound-producing action types. We propose a signal-processing based approach for segmentation, based on a tracking of the salient characteristics over time, and dually Gestalt-based segmentation decisions based on changes. Tracking of pitched sound relies on partial tracking, whereas the analysis of noisy sound requires tracking of larger frequency bands possibly varying over time. The resulting sound objects are then described based on Schaeffer’s taxonomy and morphology, expressed first in the form of numerical descriptors, each related to one type of taxonomy (percussive/sustained/iterative, stable/moving pitch vs unclear pitch) or morphology (such as grain). This multidimensional feature representation is further divided into discrete categories related to the different classes of sounds. The typological and morphological categorisation is driven by the theoretical and experimental framework of the morphodynamical theory. We first experiment on isolated sounds from the Solfège des objets sonores—which features a large variety of sound sources—before considering more complex configurations featuring a succession of sound objects without silence or with simultaneous sound objects. Analytical results are visualised in the form of graphical representations, aimed both for musicology and music pedagogy purposes. This will be applied to the graphical descriptions of and browsing within large music catalogues. The application of the analytical descriptions to music creation is also investigated.

We present an ongoing project dedicated to the transmutation of a collection of field recordings of Norwegian folk music established in the 1960s into an easily accessible online catalogue augmented with advanced music technology and computer musicology tools. We focus in particular on a major highlight of this collection: Hardanger fiddle music. The studied corpus was available as a series of 600 tape recordings, each tape containing up to 2 hours of recordings, associated with metadata indicating approximate positions of pieces of music. We first need to retrieve the individual recording associated with each tune, through the combination of an automated pre-segmentation based on sound classification and audio analysis, and a subsequent manual verification and fine-tuning of the temporal positions, using a home-made user interface. Note detection is carried out by a deep learning method. To adapt the model to Hardanger fiddle music, musicians were asked to record themselves and annotate all played note, using a dedicated interface. Data augmentation techniques have been designed to accelerate the process, in particular using alignment of varied performances of same tunes. The transcription also requires the reconstruction of the metrical structure, which is particularly challenging in this style of music. We have also collected ground-truth data, and are conceiving a computational model. The next step consists in carrying out detailed music analysis of the transcriptions, in order to reveal in particular intertextuality within the corpus. A last direction of research is aimed at designing tools to visualise each tune and the whole catalogue, both for musicologists and general public.

Kunstig intelligens kan hjelpe deg å forstå musikk bedre. UiO-forsker Olivier Lartillot jobber for at ny teknologi kan åpne folks ører for ny musikk.

Algorithms and technology have so far helped listeners to more of the same music. Now, UiO researchers are working on new technology that can get people interested in a greater musical variety.

This paper presents a Hardanger fiddle dataset “HF1” with polyphonic performances spanning five different emotional expressions: normal, angry, sad, happy, and tender. The performances thus cover the four quadrants of the activity/valence-space. The onsets and offsets, together with an associated pitch, were human-annotated for each note in each performance by the fiddle players themselves. First, they annotated the normal version. These annotations were then transferred to the expressive performances using music alignment and finally human-verified. Two separate music alignment methods based on image registration were developed for this purpose; a B-spline implementation that produces a continuous temporal transformation curve and a Demons algorithm that produces displacement matrices for time and pitch that also account for local timing variations across the pitch range. Both methods start from an “Onsetgram” of onset salience across pitch and time and perform the alignment task accurately. Various settings of the Demons algorithm were further evaluated in an ablation study. The final dataset is around 43 minutes long and consists of 19 734 notes of Hardanger fiddle music, recorded in stereo. The dataset and source code are available online. The dataset will be used in MIR research for tasks involving polyphonic transcription, score alignment, beat tracking, downbeat tracking, tempo estimation, and classification of emotional expressions.

Analyzing musical performances is a challenging and emergent field of computational music research, aiming to reveal performance patterns and link them to musical contexts. There exists a modest amount of computational research on Hardanger fiddle performances. The MIRAGE research project is currently contributing to this scientific body, developing advanced MIR frameworks that build on recent musicological research. This paper presents the development and evaluation of two Max/MSP/Jitter software applications for music analysis and data visualization that integrate contemporary research perspectives on the complex rhythmical structuring of springar performances, investigating how we can design user-friendly computational tools that explore performance patterns in Hardanger fiddle music, in collaboration with MIRAGE. Based on a small questionnaire and a few operational tests, the study shows an interest in more effective software tools capable of revealing complex interrelations between musical dimensions in Hardanger fiddle performances. Additionally, the study highlights design considerations for tools aiming to increase the availability of computational music research in the field of musicology, such as cross-compatibility and integrated features that actively facilitate nuanced interpretation processes.

MIRtoolbox is a Matlab toolbox dedicated to the analysis of music and sound from audio recordings and to the extraction of musical features such as tonality, rhythm, or structures. It has also been used for non- musical applications, such as in Non Destructive Testing, and with non-audio signals. In this issue of the newsletter, the YAN discusses the MIRtoolbox with Olivier Lartillot (RITMO Centre for Interdisciplinary Studies in Rhythm, Time and Motion, University of ̽����ѡ, Norway) and Petri Toiviainen (University of Jyväskylä, Finland) You can also check out the MIRtoolbox website at: shorturl.at/oA038

MIRtoolbox is a Matlab toolbox dedicated to the analysis of music and sound from audio recordings and to the extraction of musical features such as tonality, rhythm, or structures. It has also been used for non- musical applications, such as in Non Destructive Testing, and with non-audio signals. In this issue of the newsletter, the YAN discusses the MIRtoolbox with Olivier Lartillot (RITMO Centre for Interdisciplinary Studies in Rhythm, Time and Motion, University of ̽����ѡ, Norway) and Petri Toiviainen (University of Jyväskylä, Finland) You can also check out the MIRtoolbox website at: shorturl.at/oA038

The GrooveToolbox is a new Python toolbox implementing various algorithms, new and pre-existing, for the analysis and comparison of symbolic drum loops, including rhythm features, similarity metrics and microtiming features. As part of the GrooveToolbox we introduce two new metrics of rhythm similarity and four features for describing the significant properties of microtiming deviations in drum loops. Based on a two-part perceptual evaluation, we show these four new microtiming features can each correlate to similarity perception, and be used with rhythm similarity metrics to improve personalized similarity models for drum loops. A new measure of structural rhythmic similarity is also shown to correlate more strongly to similarity perception of drum loops than the more com- monly used Hamming distance. These results point to the potential application of the GrooveToolbox and its new features in drum loop analysis for intelligent music production tools. The GrooveToolbox may be found at: https://github.com/fredbru/GrooveToolbox

This paper develops the hypothesis that symbolic drum patterns can be represented in a reduced form as a sim- ple oscillation between two states, a Low state (commonly associated with kick drum events) and a High state (often associated with either snare drum or high hat). Both an onset time and an accent time is associated to each state. The systematic inference of the reduced form is formal- ized. This enables the specification of a rhythmic struc- tural similarity measure on drum patterns, where reduced patterns are compared through alignment. The two-state representation allows a low computational cost alignment, once the complex topological formalization is fully taken into account. A comparison with the Hamming distance, as well as similarity ratings collected from listeners on a drum loop dataset, indicates that the bistate reduction enables to convey subtle aspects that goes beyond surface-level com- parison of rhythmic textures.

We present a new system for real-time visualisation of music performance, focused for the moment on a fugue played by a string quartet. The basic principle is to offer a visual guide to better understand music using strategies that should be as engaging, accessible and effective as possible. The pitch curves related to the separate voices are drawn on a space whose temporal axis is normalised with respect to metrical positions, and aligned vertically with respect to their thematic and motivic classification. Aspects related to tonality are represented as well. We describe the underlying technologies we have developed and the technical setting. In particular, the rhythmical and structural representation of the piece relies on real-time polyphonic audio-to-score alignment using online dynamic time warping. The visualisation will be presented at a concert of the Danish String Quartet, performing the last piece of The Art of Fugue by Johann Sebastian Bach.

Computational models for the analysis of tempo, metre and the tracking of beats have made significant progress during the last decades. I first present a synthetic overview of the state of the art. Up to recently, classical approaches were based on signal processing, with the integration of heuristics based on assumptions related to music perception and cognition. The standard approach is to first detect percussive events through the establishment of an accentuation curve, followed by periodicity detection, and the construction and tracking of meter. Because rhythmic emphasis can develop on various metrical levels across time, it is necessary to track the metrical structure on multiple levels. I show the benefit of such detailed analysis with the use of a model I have developed, and which obtained one of the highest grades in the MIREX tempo estimation competition. New approaches based on deep learning have achieved impressive progress and have largely surpassed signal-processing-based approached (including mine) in the recent yearly editions of MIREX. One limitation of these approaches, at least in their current stages, is that they appear as black boxes able to imitate a particular behaviour for which they were trained on particular examples. As such, they hardly offer insight on the cognitive mechanisms underlying the perception of metre. I will discuss the limitations of signal processing approaches and highlight the complexity of the musical structure. Pulsation in music is not always expressed through a periodic repetition of percussive events, but may emerge from a subtle propagation of motivic or harmonic structures. I present an approach under development that models the different components of music analysis and combine them altogether, extending further Lerdahl and Jackendoff’s vision. Motivic repetition, which plays a core role, is also one of the dimensions that is the most difficult to model and automate.

Music video for "Aether Trouble" by PYSJ. With: Pia Watson Víctor González (RITMO*) Torgeir Koppang (PYSJ) Solveig Sørbø (PYSJ) Cinematography: John Bentham Camera Operators: Morten Malerstuen (camera and additional cinematography) Kristoffer Haugen Lighting: Nicholas Blakstad Andresen Morten Malerstuen Breath data collection (using FLOW* by SweetZpot): Víctor González, RITMO* Visualization of breath data and audio: MIRAGE* and MIRtoolbox by Olivier Lartillot, RITMO* Editing: María Isabel González Story / Concept: Solveig Sørbø Olivier Lartillot Pia Watson Directed by: John Bentham Solveig Sørbø Consultant: Nicholas Blakstad Andresen Produced by: Solveig Sørbø Extras: The dogs Vips and Willy Special thanks: Sagar Sen Naín Mendoza Fonseca Turid Svensøy SweetZpot RITMO* Workaway.info Filter Musikk Thanks to F21 for letting us use their studio MUSIC: Find the audio track on your plattform of choice: https://fanlink.to/aether Find the audio track on your plattform of choice: https://fanlink.to/aether Aether Trouble by PYSJ Written by: Solveig Sørbø Performed by PYSJ Solveig Sørbø Torgeir Koppang Andreas Rødland Haga Stig Frogner Mixing: Stig Frogner Mastering: Björn Engelmann / The Cutting Room Production: Solveig Sørbø * RITMO Centre for Interdisciplinary Studies in Rhythm, Time and Motion, University of ̽����ѡ. This work was partially supported by the Research Council of Norway through its Centres of Excellence scheme, project number 262762. * Learn more about the visualization technologies on the MIRAGE project website: http://bit.ly/MirageProject * Data was collected using FLOW™ sensors from SweetZpot. For more information about this technology see: https://www.sweetzpot.com/flow

During this presentation, Dr. Olivier Lartillot will give an overview of MIRtoolbox - a Matlab application enabling the extraction of a large range of audio and musical descriptions from recordings. MIRtoolbox is designed to be easy to use both for teaching at any level and for advanced research in musicology, signal analysis and music cognition. One initial aim of MIRtoolbox was to study the relationship between musical features and emotions evoked by music. MIRtoolbox focuses on signal-processing-based approaches that offer limited understanding of music. Lartillot is also developing computational methods for the analysis of notated music, starting from motivic analysis and aiming at building a comprehensive framework where audio and score are combined together.

In the first part of this talk, I will give a short and broad overview of the MiningSuite, a Matlab toolbox that combines sound, music and gesture analysis. I will give a quick tour of the various types of sound and music analyses that can be carried out using the toolbox, covering a large range of musical dimensions such as timbre, rhythm, harmony or structure. The MiningSuite, integrating previous toolboxes such as MIRtoolbox and MIDI toolbox, can be used both for the analysis of audio recordings and of “symbolic” representations such as MIDI files. I will also present the current integration of motion capture and gesture analysis (from the MoCap Toolbox), as well as other sensor data such as breathing. The benefit of articulating these different types of analyses into a single framework will be demonstrated. In the second part, I will present a project aimed at automatically extracting melodic gestures from a catalogue of folk music recordings from the National Library of Norway. While melodic lines can be easily extracted from a cappella songs, the task is more challenging for other types of music such as Hardanger fiddle music. In such cases, we need to automatically transcribe the recordings and track melodic voices throughout the counterpoint of each composition. I will also present an iPhone app that enables to draw a gesture in the air with the phone and to find pieces of music from the catalogue that is characterised by a similar musical gesture.

AcousMuScope is a new software for music analysis of audio recordings, focusing on the graphical interface to browse into the analyses.