Nádia Sousa Carvalho

Abstract
This paper introduces Motiv, a dataset of expert saxophonist recordings illustrating parallel, similar, oblique, and contrary motions. These motions are variations of three phrases from Jesús Villa-Rojo's "Lamento,"with controlled similarities. The dataset includes 116 audio samples recorded by four tenor saxophonists, each annotated with descriptions of motions, musical scores, and latent space vectors generated using the VocalSet RAVE model. Motiv enables the analysis of motion types and their geometric relationships in latent spaces. Our preliminary dataset analysis shows that parallel motions align closely with original phrases, while contrary motions exhibit the largest deviations, and oblique motions show mixed patterns. The dataset also highlights the impact of individual performer nuances. Motiv supports a variety of music information retrieval (MIR) tasks, including gesture-based recognition, performance analysis, and motion-driven retrieval. It also provides insights into the relationship between human motion and music, contributing to real-time music interaction and automated performance systems. © 2025 Copyright held by the owner/author(s).

Abstract
This article investigates sampling strategies in latent space navigation to enhance co-creative music systems, focusing on timbre latent spaces. Adopting Villa-Rojo's 'Lamento' for tenor saxophone and tape as a case study, we conducted two experiments. The first assessed traditional corpus-based concatenative synthesis sampling within the RAVE model's latent space, finding that sampling strategies gradually deviate from a given target sonority while still relating to the original morphology. The second experiment aims at defining sampling strategies for creating variations of an input signal, namely parallel, contrary, and oblique motions. The findings expose the need to explore individual model layers and the geometric transformation nature of the contrary and oblique motions that tend to dilate the original shape. The findings highlight the potential of motion-aware sampling for more contextually aware and expressive control of music structures via CBCS.

Abstract
This paper presents a comprehensive investigation into the explainability and creative affordances derived from navigating a latent space generated by Realtime Audio Variational AutoEncoder (RAVE) models. We delve into the intricate layers of the RAVE model's encoder and decoder outputs by leveraging a novel timbre latent space that captures micro-timbral variations from a wide range of saxophone extended techniques. Our analysis dissects each layer's output independently, shedding light on the distinct transformations and representations occurring at different stages of the encoding and decoding processes and their sensitivity to a spectrum of low-to-high-level musical attributes. Remarkably, our findings reveal consistent patterns across various models, with the first layer consistently capturing changes in dynamics while remaining insensitive to pitch or register alterations. By meticulously examining and comparing layer outputs, we elucidate the underlying mechanisms governing saxophone timbre representation within the RAVE framework. These insights not only deepen our understanding of neural network behavior but also offer valuable contributions to the broader fields of music informatics and audio signal processing, ultimately enhancing the degree of transparency and control in co-creative practices within deep learning music frameworks. © 2024. This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 Unported License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original.

Abstract
This paper investigates sampling strategies within latent spaces for music generation, focusing on (chordified) J.S. Bach Chorales and utilizing MusicVAE as the generative model. We conduct an experiment comparing three sampling and interpolation strategies within the latent space to generate chord progressions - from a discrete vocabulary of Bach's chords - to Bach's original chord sequences. Given a three-chord sequence from an original Bach chorale, we assess sampling strategies for replacing the middle chord. In detail, we adopt the following sampling strategies: (1) traditional linear interpolation, (2) k-nearest neighbors, and (3) k-nearest neighbors combined with angular alignment. The study evaluates their alignment with music theory principles of functional harmony embedding and voice-leading to mirror Bach's original chord sequences. Preliminary findings suggest that knearest neighbors and k-nearest neighbors combined with angular alignment closely align with the tonal function of the original chord, with k-nearest neighbors excelling in bass line interpolation and the combined strategy potentially enhancing voice-leading in upper voices. Linear interpolation maintains aspects of voice-leading but confines selections within defined tonal spaces, reflecting the nonlinear characteristics of the original sequences. Our study contributes to the dynamics of latent space sampling for music generation, offering potential avenues for enhancing explainable creative strategies. © 2024. This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 Unported License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original.

Abstract
Extensive computational research has been dedicated to detecting keys and modes in tonal Western music within the major and minor modes. Little research has been dedicated to other modes and musical expressions, such as folk or non-Western music. This paper tackles this limitation by comparing traditional template-based with unsupervised machine-learning methods for diatonic mode detection within folk music. Template-based methods are grounded in music theory and cognition and use predefined profiles from which we compare a musical piece. Unsupervised machine learning autonomously discovers patterns embedded in the data. As a case study, the authors apply the methods to a dataset of Irish folk music called The Session on four diatonic modes: Ionian, Dorian, Mixolydian, and Aeolian. Our evaluation assesses the performance of template-based and unsupervised methods, reaching an average accuracy of about 80%. We discuss the applicability of the methods, namely the potential of unsupervised learning to process unknown musical sources beyond modes with predefined templates.