Personalizing Automatic Speech Recognition Models for Non-normative Speech using MoE

Automatic Speech Recognition (ASR) for individuals with impaired speech remains a significant challenge due to extreme data scarcity and high acoustic variability. This project builds upon a successful, data-efficient Bayesian personalization framework (Variational Inference Low-Rank Adaptation, VI-LoRA), coming from our team. This project aims to further increase personalised ASR performance by exploring the implementation of Hydra VI-LoRA. Hydra VI- LoRA, is a novel Mixture-of-Experts (MoE) architecture capable of learning specialized adapters for different speakers or phonetic challenges within a single model. This project is designed for a highly motivated and independent student eager to take ownership of a cutting-edge research topic.