Multi-Task and Transfer Learning in Low-Resource Speech Recognition

Abstract

This thesis investigates methods for Acoustic Modeling in Automatic Speech Recog- nition, assuming limited access to training data in the target domain. The Acoustic Models of interest are Deep Neural Network Acoustic Models (in both the Hybrid and End-to-End approaches), and the target domains in question are either different languages or different speakers. Inductive bias is transfered from a source domain during training, via Multi-Task Learning or Transfer Learning. With regards to Multi-Task Learning, Chapter (5) presents experiments which explicitly incorporate linguistic knowledge (i.e. phonetics and phonology) into an auxiliary task during neural Acoustic Model training. In Chapter (6), I investigate Multi-Task methods which do not rely on expert knowledge (linguistic or otherwise), by re-using existing parts of the Hybrid training pipeline. In Chapter (7), new tasks are discovered using unsupervised learning. In Chapter (8), using the “copy-paste” Transfer Learning approach, I demonstrate that with an appropriate early-stopping criteria, cross-lingual transfer is possible to both large and small target datasets. The methods and intuitions which rely on linguistic knowledge are of interest to the Speech Recognition practitioner working in low-resource domains. These same sections may be of interest to the theoretical linguist, as a study of the relative import of phonetic categories in classification. To the Machine Learning practitioner, I hope to offer approaches which can be easily ported over to other classification tasks. To the Machine Learning researcher, I hope to inspire new ideas on addressing the small data problem.