Computer Science - Post-Training Adaptation of Generative Audio Models

Supervisor: Dr Jacopo de Berardinis

Supervisor bio: Dr. Jacopo de Berardinis is an interdisciplinary Lecturer in Computer Science at the University of Liverpool, where he develops human-centric AI systems designed to support and safeguard human creativity and musical heritage. His research expertise is at the intersection of Knowledge Engineering, Machine Learning and Music Informatics. He focuses on building fair, interoperable, and responsible systems within the complex socio-technical ecosystem of the creative industries and musical heritage.

Read more about my research & projects at http://www.jacopodeberardinis.com

Email: jacodb@liverpool.ac.uk

School: School of Computer Science and Informatics 

Department: Computer Science

Module code: COMP298

Suitable for students of Computer Science, Engineering 

Desirable experience/requirements: 

Knowledge and prior experience with training machine learning models using frameworks such as PyTorch, JAX, or Tensorflow (expected).

Experience with processing sequential data, particularly audio, is desirable.

Fundamentals of Generative AI models is also desirable.

Places available: 2

Start date: 15th June 2026

Project length: 8 weeks 

Virtual option: No

Hybrid option: Yes

Project description: 

The rapid emergence of AI music generators like MusicGen, Stable Diffusion, and AudioLDM has unlocked new creative frontiers but has also ignited critical debates around copyright, mimicry, and artist rights. As these powerful models are often trained on indiscriminate datasets, the creative industry faces a significant challenge: ensuring AI respects intellectual property without stifling innovation. This project addresses this urgent gap by developing technical solutions for "Responsible AI", focusing on how we can engineer models that are not only high-quality but also compliant and ethically aligned with the creative community.

This research focuses on the advanced manipulation of generative deep learning models to solve these alignment issues. The selected student will investigate novel algorithmic interventions to "steer" pre-trained models, exploring methods to selectively manage the reproduction of specific musical repertoires or styles. Rather than relying on simple output filtering, this project dives into the model architecture itself, aiming to adapt the latent space or weights of large-scale audio models. The goal is to demonstrate that it is possible to constrain a generative model’s access to specific protected data distributions while preserving its overall creative capabilities.

This project presents a unique opportunity for students to delve into the intersection of Deep Learning and Music Information Retrieval (MIR). Proficiency in Python and practical experience with PyTorch are essential requirements, as the intern will be adapting complex neural network architectures. While background knowledge in audio signal processing is beneficial, the primary focus is on deep learning engineering. This internship offers a unique trajectory toward academic publication; depending on the project's progress, the student will have the opportunity to work towards a paper submission to leading conferences such as ICASSP or ISMIR. Prospective applicants are highly encouraged to make informal inquiries about the project prior to submitting their applications.

Additional requirements: N/A