Session 4 — Paper Presentations

Musical Pattern Discovery at Regular Time Intervals Through Mathematical Morphology: An Application to Metric Structure Analysis

Mathys Daniel, Paul Lascabettes, Moreno Andreatta and Isabelle Bloch

Keywords: Pattern Discovery, Symbolic Music Analysis, Computational Musicology, Mathematical Morphology

Abstract

Musical pattern discovery is central in computational music analysis, as repeated structures reveal information about compositional intentions and listener perceptions. Traditional symbolic pattern discovery methods usually extract all repeating patterns in musical data but sometimes discover many patterns that are musically irrelevant. To overcome these limitations, we introduce in this paper a method based on mathematical morphology that specifically targets patterns repeated at exact regular time intervals. By analyzing the length of the extracted patterns, our approach provides insights into the rhythmic signature of a piece and enables automatic structural analysis. We apply this method to Polygondwanaland by King Gizzard & The Lizard Wizard to describe its rhythmic complexity and extensive use of polymetric structures. The proposed approach extracts fewer patterns than state-of-the-art algorithms while significantly reducing computation time. These results highlight both the analytical potential of the method for musicological study and its effectiveness as a computational tool for large-scale pattern discovery.

 

Diachronic Modeling of Tonal Coherence on the Tonnetz Across Classical and Popular Repertoires

Weilun Xu, Edward Hall and Martin Rohrmeier

Keywords: tonal coherence, Tonnetz, computational musicology, tonal diffusion model, cross-genre comparison, pitch-class distributions

Abstract

How do different musical traditions achieve tonal coherence? Most computational measures to date have analysed tonal coherence in terms of a single dimension, whereas a multi-dimensional analyses have not been sufficiently explored. We propose a new model drawing on the concept of the Tonnetz - we define two partially independent measures: \emph{tonal focus}, the concentration of pitch content near a tonal center; and \emph{tonal connection}, the degree to which pitch content reflects structured intervallic pathways back to that center. Analyzing over 2,800 pieces from Western classical and popular traditions, we find that these traditions occupy overlapping yet distinguishable regions of the two-dimensional space. Popular music shows higher tonal focus, while classical music exhibits higher tonal connection. Our complementary measures ground the differences between different tonal styles in quantitative evidence, and offer interpretable dimensions for computational music analysis and controllable generation.

 

Mining Contrast Patterns in Jazz Guitar Solos

Arne Eigenfeldt and Darrell Conklin

Keywords: computational musicology, jazz guitar, improvisation, analysis

Abstract

Computational musicology is the part of musicology that employs computational methods for the analysis of music corpora. This paper uses a pattern discovery method to investigate the melodic patterns employed by five jazz guitarists, discovering those that are overrepresented in their solos. The paper also compares the patterns of these guitarists with transcriptions of jazz solos of other instruments to determine whether there are particular improvisatory patterns that are over-represented on the guitar.

 

Music Similarity Estimation from Jukebox Embeddings

Johannes Grafinger, Gerhard Widmer and Jan Schlüter

Keywords: Music Similarity Estimation, Generative Models, Pretrained Models

Abstract

Content-based music similarity estimation aims to provide a measure of similarity between music audio recordings, disregarding any associated metadata. This has potential applications in e.g. cold-start music recommendation scenarios, where there is no user interaction data. In this work, we propose JukeSim, a new approach to estimating musical similarity based on embeddings created by Jukebox, a generative music model. In evaluations on multiple datasets, we find that JukeSim performs on par with an existing learning-based approach, and both outperform the best hand-designed approaches, which had never been compared to recent work before. Even though the required compute for JukeSim is higher than for other measures, we demonstrate the viability of utilizing generative embeddings for this task, encouraging further research.

 

 

 

 

 

 

Unsupervised Evaluation of Deep Audio Embeddings for Music Structure Analysis

Axel Marmoret

Keywords: Music Structure Analysis, Deep Audio Models, Unsupervised Learning

Abstract

Music Structure Analysis (MSA) aims to uncover the high-level organization of musical pieces. State-of-the-art methods are often based on supervised deep learning, but these methods are bottlenecked by the need for heavily annotated data and inherent structural ambiguities. In this paper, we propose an unsupervised evaluation of nine open-source, generic pre-trained deep audio models, on MSA. For each model, we extract barwise embeddings and segment them using three unsupervised segmentation algorithms (Foote's checkerboard kernels, spectral clustering, and Correlation Block-Matching (CBM)), focusing exclusively on boundary retrieval. Our results demonstrate that modern, generic deep embeddings generally outperform traditional spectrogram-based baselines, but not systematically. Furthermore, our unsupervised boundary estimation methodology generally yields stronger performance than recent linear probing baselines. Among the evaluated techniques, the CBM algorithm consistently emerges as the most effective downstream segmentation method. Finally, we highlight the artificial inflation of standard evaluation metrics and advocate for the systematic adoption of “trimming'', or even “double trimming'' annotations to establish more rigorous MSA evaluation standards.

 

Adapting Deep Learning-based Meter Tracking to Carnatic Music

Satyajeet Prabhu, Thomas Nuttall, Vivek Vijayan and Martín Rocamora

Keywords: Computational Musicology, Carnatic Music, Indian Art Music, Rhythm Analysis, Beat Tracking, Downbeat Tracking, Deep Learning

Abstract

Beat and downbeat tracking, jointly referred to as meter tracking, is a fundamental task in the computational analysis of musical rhythm. However, state-of-the-art models often perform less reliably on musical traditions underrepresented in their training data. The South Indian musical tradition known as Carnatic music is one such case. In this study, we evaluate two state-of-the-art meter-tracking models–the Temporal Convolutional Network and Beat This!–in the Carnatic music context, and explore musically informed strategies to adapt them to this tradition.
 Our results show that while off-the-shelf models do not always outperform the existing baseline for Carnatic music, their performance improves substantially with domain-specific training or transfer learning, matching or surpassing the existing state-of-the-art, and demonstrating that deep learning models can be effectively adapted to underrepresented traditions. Our code is made publicly available, and the Carnatic meter-tracking models are released as part of an existing tools repository for the computational analysis of Carnatic music.

 

 

Score-Informed Transformer for Refining MIDI Velocity in Automatic Piano Transcription

Zhanhong He, Roberto Togneri and David Huang

Keywords: Automatic Music Transcription, MIDI Velocity Estimation, Piano Transcription, Score-informed, Transformer

Abstract

MIDI velocity is crucial for capturing expressive dynamics in human performances. In practical scenarios, a music score with inaccurate velocities may be available alongside the performance audio (e.g., music education and free online archives), enabling the task of score-informed MIDI velocity estimation. In this work, we propose a modular, lightweight score-informed Transformer correction module that refines the velocity estimates of Automatic Music Transcription (AMT) systems. Leveraging high-quality expressive datasets for rigorous evaluation, we focus on piano transcription, integrating our module into multiple existing systems. Trained exclusively on the MAESTRO dataset, our method consistently reduces velocity errors and improves cross-dataset generalization to SMD and MAPS datasets. Under this training protocol, integrating our score-informed module with HPT (named Score-HPT) establishes a new state-of-the-art, outperforming existing score-informed methods and piano transcription models while adding only 1 M parameters. Furthermore, our framework remains robust to temporal misalignments caused by imperfect audio-score synchronization, a common real-world challenge.

 

Reharmonization Grammar Rules in Jazz Improvisation

George Konstandelakis and Christina Anagnostopoulou

Keywords: jazz improvisation, reharmonization, harmonic analysis, computational musicology, generative grammar

Abstract

Reharmonization is a core mechanism through which jazz improvisers shape cadential drive, local tension, and harmonic direction. This paper proposes an interpretable grammar of jazz reharmonization consisting of 10 generative rules (19 including subrules), derived from a corpus of 30 piano improvisations using left-hand voicings as evidence of harmonic intent and informed by established accounts of harmonic function and jazz reharmonization. The grammar formalizes recurrent mechanisms including quality reinterpretation, tritone substitution, cadential insertion and omission, diminished interpolation, and chromatic approach. Evaluation suggests that reharmonization is widespread and frequently compositional. In the 30-improvisation source corpus, 50.5\% of 192 annotated occurrences require multi-rule primary explanations, and 25.5\% admit a second annotated explanation. In an external validation on 30 additional improvisations, the fixed rule inventory achieves full coverage of 350/350 under the present annotation protocol, with 60.6\% of occurrences requiring multi-rule explanations and 26.0\% admitting a second annotated explanation. These results suggest that the rule inventory transfers beyond the source corpus. Implemented in an explainable prototype (“Reharmonizer''), the transformation grammar generates alternative reharmonizations with explicit derivation traces. The study thus extends grammar-based harmonic modeling to improvisational reharmonization and offers a compact rule inventory for analysis, computation, and pedagogy.

A Computational Model for Predicting Perceived Tension in Electronic Dance Music Buildups

Stefanos Biliousis and Sofia Dahl

Keywords: Musical tension modeling, Electronic Dance Music (EDM), Music Information Retrieval, Gestalt principles, Audio feature extraction, Pitch proximity, Spectral centroid, Rhythmic density, Computational music analysis

Abstract

This study presents a computational model of perceived musical tension in Electronic Dance Music (EDM) buildup sections, grounded in the Gestalt principles of pitch proximity, timbral similarity, and temporal proximity. The model derives a composite Tension Index from audio features representing pitch, timbre, and temporal density. We evaluated the model in a continuous-rating study in which 14 listeners rated twelve extended excerpts spanning breakdown, build-up, and drop. Continuous listener ratings were compared with the model using Pearson correlation. Accounting for a 2.0 s response lag, the model correlated positively with listener ratings (mean ???? = .57, 95% bootstrap CI [.51, .62], ???? < .001). Inter-rater reliability, assessed per track, was high (ICC(2,????) .85 to .95). After removing the shared linear trend, the model retained significant predictive power, although a linear-ramp baseline performed better on average, indicating that substantial track-specific tension remains unexplained. These findings show that Gestalt-based audio features capture a meaningful component of perceived tension in EDM break routines and establish a quantified benchmark for future genre-aware tension models.

 

BMdataset: A Musicologically Curated LilyPond Dataset

Matteo Spanio, Ilay Guler and Antonio Rodà

Keywords: dataset, symbolic music, lilypond, continuous pretraining, finetuning, codebert

Abstract

Symbolic music research has relied almost exclusively on MIDI-based datasets; text-based engraving formats such as LilyPond remain unexplored for music understanding. We present BMdataset, a musicologically curated dataset of 347 LilyPond scores (2,646 movements) transcribed by experts directly from original Baroque manuscripts, with metadata covering composer, musical form, instrumentation, and sectional attributes. Building on this resource, we introduce LilyBERT (https://huggingface.co/csc-unipd/lilybert), a CodeBERT-based encoder adapted to symbolic music through vocabulary extension with 115 LilyPond-specific tokens and masked language model pre-training. Linear probing on the out-of-domain Mutopia corpus shows that, despite its modest size (~90M tokens), fine-tuning on BMdataset alone outperforms continuous pre-training on the full PDMX corpus (~15B tokens) for both composer and style classification, demonstrating that small, expertly curated datasets can be more effective than large, noisy corpora for music understanding. Combining broad pre-training with domain-specific fine-tuning yields the best results overall (84.3% composer accuracy), confirming that the two data regimes are complementary. We release the dataset, tokenizer, and model to establish a baseline for representation learning on LilyPond.