Factorized adapters for robust domain adaptation in ASR

Automatic Speech Recognition (ASR) performance is readily degraded by environmental noise, reverberation, and other forms of distortion. A significant challenge in noise robust ASR is dealing with forms of environmental noise that were unseen during model training, especially when multiple types of distortion are present simultaneously. In this thesis, we propose the use of factorized adapters for robust ASR. By utilizing a framework that enables us to compose multiple adapters that have each been trained to model one type of noise, we demonstrate that our models can generalize in a zero-shot fashion to distortion combinations that were unseen during training.