Paper Session 6b: AI & Machine Learning

Giovanni Roma and Alba Francesca Battista: “Supervised Memory: How Machines Can Preserve What We Cannot Hold”

This paper presents an AI framework for preserving electroacoustic works threatened by technological obsolescence and vanishing performance knowledge. Through supervised annotation as “composition of comprehension,” we transform machine learning into active interpretation rather than passive archiving. Our approach employs a two-level vocabulary system distinguishing universal from composer-specific notational elements, enabling systematic knowledge transfer across diverse repertoires. We ground the framework in one implemented reconstruction—Jonathan Harvey’s Ricercare una melodia from incomplete documentation—and outline two further experimental
fronts: analyzing context-dependent notation in Stockhausen’s Solo, and exploring annotation possibilities in Boulez’s spatial coordinates. The methodology treats annotation not as neutral transcription but as interpretive translation, where each label embeds aesthetic decisions and performance practice. Harvey’s implementation revealed how editorial simplification between 1984 and 2003 editions created cascading performance challenges, validating our recovery of embedded procedural knowledge. The framework progresses from mechanical reproduction through systematic reading to conscious reactivation, establishing foundations for computational preservation while acknowledging fundamental limits. We argue that effective preservation requires not static archives but living traditions maintained by transparent, contestable machine interpretations. This positions AI-based complements as participants in musical preservation rather than mere repositories, preserving both structural relationships and the reasoning patterns that animate them.

Abhirup Saha, Hans-Ulrich Berendes, Meinard Müller, and Ben Maman: “Snapping Matters: Context-Aware Onset Refinement for Automatic Music Transcription”

Precise note-level annotations are critical for training automatic music transcription (AMT) systems, in particular note-onset labels, which form a core component of many recent AMT systems. However, high-quality annotations for real-world recordings are scarce. Sequence-level score–audio alignment methods such as dynamic time warping provide only coarse correspondence, making a local refinement step necessary. This refinement step, known as snapping, adjusts aligned score onsets using peaks in a neural onset posteriorgram and often determines whether weakly aligned score–audio pairs become usable training data at all. Despite its practical importance, snapping is typically treated as a simple post-processing heuristic and implemented with greedy local decisions. We present a systematic analysis of snapping strategies for training instrument-agnostic transcribers, demonstrating that snapping is essential for learning from weakly aligned data. Building on this, we formulate snapping as a per-pitch assignment problem and solve it via bipartite graph matching, yielding context-aware onset decisions under overlapping refinement windows and uncertain initial alignments. Extensive cross-dataset experiments across piano, chamber, and orchestral recordings show improved onset alignment and transcription accuracy over greedy snapping, with gains increasing for wider snapping windows and coarser initial alignments. Qualitative examples are provided on our project page: https://abhirupsaha8.github.io

Yu Foon Darin Chau and Andrew Horner: “Classical Music Mashup System and Compatibility Heuristics”

We investigate symbolic classical music mashups and introduce a retrieval-based pipeline for generating them. Unlike audio-domain mashups, symbolic mashups offer perfect voice isolation and allow for post-generation reinterpretation of tempo, dynamics, and instrumentation. While prior work in audio mashups emphasises harmony, rhythm, and balance, symbolic mashups in classical repertoires remain underexplored and lack clear compatibility heuristics. To this end, we conduct controlled listening tests on classical music excerpts to isolate factors shaping perceived compatibility. Results indicate effective mashups should respect the recognizability of motivic materials, underlying cadential logic, and be presented polyphonically. We designed a symbolic mashup pipeline for classical piano music around these findings that maximises pairwise piece compatibility. We discuss implications and limitations for algorithmic composition, pedagogical tools, and future extensions to broader styles, longer forms, and richer evaluative methodologies.