« All Events

Paper Session 5b: AI, Machine Learning & Pedagogy

May 13 @ 9:00 am - 10:30 am
Three papers will be presented and discussed:

 

Jeff Kaiser and Gregory Taylor: “Building Loopers: A Pedagogical Framework for Teaching Creative Software Design Through Iterative Tool Construction in Max, gen~, and RNBO”

This paper introduces the ideas behind our open-access project “Building Live Loopers in Max.”1 The project presents a hybrid pedagogical and technical framework in which students learn signal processing concepts by constructing live-looping tools in Max, gen~, and RNBO. By engaging with buffer operations, timing structures, playback manipulation, and parameter mapping, students develop technical fluency and musical understanding simultaneously. We introduce a sequence of modular, step-by-step looper designs, a color-coded instructional method for visualizing patcher development, and a cross-environment workflow that reinforces transferable pro-
gramming habits. Our coursework is designed to be sufficiently open-ended that students, while grounded in familiar musical contexts, are encouraged to exercise curiosity and explore creative directions beyond the methods presented. Drawing on Dehaene’s work on curiosity and Eagle-
man’s writing on relevance, the design aims to engage intrinsic motivation and support students in forming novel connections and actively experimenting with musical ideas. This approach positions looper construction as a bridge between creative music-making and computational thinking, supporting both performance and pedagogical outcomes.

Nicolas Brochec and Jean-Louis Giavitto: “Automatic Following of Flute Playing Techniques for Real-Time Mixed Music: A Case Study with Antescofo and ipt~”
This paper investigates how real-time recognition of instrumental playing techniques can extend automatic score following beyond the limits of pitch-based alignment. While systems such as Antescofo provide robust and largely plug-and-play score following, their listening model is primarily designed for stable, pitched events aligned with a fixed symbolic score. This makes them difficult to adapt to extended techniques, unpitched sounds, and musical forms involving partial improvisation or open notation. To address these limitations, we explore a hybrid approach that combines multiple listening machines with complementary capabilities and allows dynamic switching between them during performance according to the musical context. Specifically, we integrate Antescofo with ipt˜, a real-time playing technique recognition system based on lightweight machine learning models. We focus on the integration of real-time instrumental playing technique recognition as a means to enrich the listening process and support technique-aware navigation of the score. We evaluate this approach on the case of extended flute techniques, assessing both the feasibility of technique aware following and the trade-off between system generality and performance. Results suggest that learning-based listening modules provide a practical compromise: they improve
robustness for specific techniques while preserving much of the plug-and-play character supporting multiple works and performers. The results highlight a promising balance between generality, specificity, and performative robustness.
Minami Kojima, Takayuki Itoh and Rafael Ramirez: “Factor Analysis of Similarity in the Same Orchestral Piece”

There have been conventional research comparing multiple performances of the same classical music piece, where many of them primarily discussed the similarity between performances but rarely delved into why those performances are similar. Furthermore, conventional similarity analysis has dealt with the similarity of single features in classical music, such as timbre, tempo, and loudness. However, since actual audiences perceive these features compounded without separating them, relying solely on single features is insufficient to fully represent musical style. To overcome this challenge, our study uses a deep learning model (VGGish, specialized in audio feature extraction) that captures high-level timbral and textural features, in addition to acoustic features (timbre, loudness, tempo), targeting the same orchestral piece. Subsequently, based on those features, we define five grouping criteria: (1) same orchestra, (2) same conductor, (3) same country of orchestra, (4) same nationality of conductor, and (5) same teacher of the conductor. We then evaluate the clus-
tering performance for each criterion. A group exhibiting high clustering performance suggests mutual similarity among the performances within that group, leading us to conclude that the corresponding criterion represents a major factor influencing performance style. The results in
this paper show that conductor identity consistently yields the strongest clustering for tempo-related features, while orchestra identity dominates timbral similarity in specific movements. We demonstrate that metadata-driven factors explain similarity beyond purely perceptual or affective similarity measures.

 

Details

Organizer

Venue