Efficient Knowledge Transfer and Adaptation for Speech and Beyond

This thesis advances the field of efficient knowledge transfer and adaptation in the realm of speech processing. It is structured to address the limitations of transfer learning in dynamically evolving audio and speech processing contexts, particularly through novel approaches for class-incremental learning, parameter-efficient adaptation, and multimodal modeling. First, we provide a comprehensive framework for class-incremental spoken language understanding, allowing models to incrementally learn new intents and entities while retaining previously acquired knowledge. Using knowledge distillation and rehearsal-based strategies, we enhance robustness against catastrophic forgetting, a key limitation in continual learning. We also introduce a unique approach to class-incremental audio classification, utilizing mutual information optimization. Second, given the prohibitive computational costs of traditional model adaptation (i.e., full fine-tuning), this thesis introduces a comprehensive framework for parameter-efficient fine-tuning of audio and speech foundation models. Furthermore, we propose new adapter designs to achieve effective transfer learning with minimal computational overhead. Finally, extending beyond unimodal settings, we propose Llama-AVSR, a new multimodal large language model with strong audio-visual speech recognition (AVSR) abilities. Llama-AVSR leverages pre-trained audio and video encoders along with a large language model to achieve state-of-the-art accuracy on the major AVSR benchmark. Notably, this model adapts pre-trained language models to the multimodal AVSR domain while keeping the core model parameters frozen, ensuring computational efficiency. Overall, this thesis provides a comprehensive framework and empirical findings that advance the application of continual learning, efficient fine-tuning, and multimodal models for speech processing tasks. These contributions pave the way for adaptive, resource-efficient speech understanding systems.

Grantor dc:publisher

Università degli studi di Trento

Year dc:date

2025

Author dc:creator

• Cappellazzo, Umberto

• Transfer Learning, Class-Incremental Continual Learning, Spoken Language Understanding, Parameter-Efficient Transfer Learning, MultiModal Large Language Models, Audio-Visual Speech Recognition

Statement dc:rights

• info:eu-repo/semantics/openAccess
• license:Creative commons
• license uri:http://creativecommons.org/licenses/by-nd/4.0/

Language dc:language

eng

Identifier

http://dx.doi.org/10.15168/11572_441850
10.15168/11572_441850

OAI identifier oai:identifier

oai:iris.unitn.it:11572/441850