Over the past decades much research has been invested in automatic speech processing, and current speech processing systems often achieve remarkably low error rates. However, systems are still dependent on matched training and test corpora, and their performance deteriorates rapidly when faced with test conditions that differ from the training data. In the machine learning community, many novel algorithms have recently been proposed to better incorporate information from unlabeled test data into the learning and classification process. Most of these techniques scale poorly to large tasks and thus seem unsuitable for speech processing. This project investigates the applicability of one recently developed paradigm, graph-based learning (GBL), for automatic speech recognition. In particular, it uses GBL as an adaptation framework where acoustic classifiers are constrained to yield outputs that vary smoothly along underlying data manifolds characterizing the test data. The project addresses the core problems of poor scalability and computational complexity of GBL by using data clustering and subselection techniques as well as sparse matrix computation. Various ways of incorporating domain information (e.g. temporal or higher-level prosodic information) into graph construction are evaluated, including the use of multiple, combined similarity measures for graph construction. Finally, online, incremental GBL techniques are being investigated in order to facilitate real-time processing. The outcome of this project is expected to significantly improve the accuracy and robustness of speech processing systems. Furthermore, it will contribute to improving the scalability of semi-supervised learning algorithms in general and thus facilitate their application to related pattern recognition fields.
