Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation

As symbolic music generation evolves, research interest is shifting toward more controlled and steerable generative processes to support creative decisions. Previous methods focus on global conditioning or fine-grained control through input sequences but often limit flexibility for real-time interve...

Full description

Saved in:
Bibliographic Details
Main Authors: Rafik Hachana, Bader Rasheed
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10879291/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849720265875914752
author Rafik Hachana
Bader Rasheed
author_facet Rafik Hachana
Bader Rasheed
author_sort Rafik Hachana
collection DOAJ
description As symbolic music generation evolves, research interest is shifting toward more controlled and steerable generative processes to support creative decisions. Previous methods focus on global conditioning or fine-grained control through input sequences but often limit flexibility for real-time interventions and require modifications to the model’s architecture. We introduce a novel symbolic music generation framework by combining a Transformer encoder-decoder with probe models, which enable us to interpret the encoder hidden state using pre-defined non-differentiable musical features, and subsequently manipulate the hidden state to achieve a set of desired attributes in the generated music. This method allows fine-grained control over specific musical features without altering the underlying model architecture. Probes can be trained jointly with the generative model or applied post-training, enabling adaptable control without retraining the model. Our experiments demonstrate that this intervention effectively influences the model output without hindering the music quality. This approach enhances both the flexibility and interpretability of symbolic music generation, enabling better real-world applicability for music generation models.
format Article
id doaj-art-69ae3eafae244937b95fd2b49b058ea7
institution DOAJ
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-69ae3eafae244937b95fd2b49b058ea72025-08-20T03:11:58ZengIEEEIEEE Access2169-35362025-01-0113280592807010.1109/ACCESS.2025.354054310879291Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music GenerationRafik Hachana0https://orcid.org/0009-0003-0651-2445Bader Rasheed1https://orcid.org/0000-0003-3874-0883Research Center for Artificial Intelligence, Innopolis University, Innopolis, RussiaResearch Center for Artificial Intelligence, Innopolis University, Innopolis, RussiaAs symbolic music generation evolves, research interest is shifting toward more controlled and steerable generative processes to support creative decisions. Previous methods focus on global conditioning or fine-grained control through input sequences but often limit flexibility for real-time interventions and require modifications to the model’s architecture. We introduce a novel symbolic music generation framework by combining a Transformer encoder-decoder with probe models, which enable us to interpret the encoder hidden state using pre-defined non-differentiable musical features, and subsequently manipulate the hidden state to achieve a set of desired attributes in the generated music. This method allows fine-grained control over specific musical features without altering the underlying model architecture. Probes can be trained jointly with the generative model or applied post-training, enabling adaptable control without retraining the model. Our experiments demonstrate that this intervention effectively influences the model output without hindering the music quality. This approach enhances both the flexibility and interpretability of symbolic music generation, enabling better real-world applicability for music generation models.https://ieeexplore.ieee.org/document/10879291/Machine learningmusic generationsymbolic musicgenerative AIprobesconditional generative models
spellingShingle Rafik Hachana
Bader Rasheed
Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
IEEE Access
Machine learning
music generation
symbolic music
generative AI
probes
conditional generative models
title Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
title_full Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
title_fullStr Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
title_full_unstemmed Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
title_short Probe-Assisted Fine-Grained Control for Non-Differentiable Features in Symbolic Music Generation
title_sort probe assisted fine grained control for non differentiable features in symbolic music generation
topic Machine learning
music generation
symbolic music
generative AI
probes
conditional generative models
url https://ieeexplore.ieee.org/document/10879291/
work_keys_str_mv AT rafikhachana probeassistedfinegrainedcontrolfornondifferentiablefeaturesinsymbolicmusicgeneration
AT baderrasheed probeassistedfinegrainedcontrolfornondifferentiablefeaturesinsymbolicmusicgeneration