Topics in Speech Recognition

title: Topics in Speech Recognition
author: D. A. Liauw Kie Fa
published in: July 2006
appeared as: Master of Science thesis
Man-machine interaction group
Delft University of Technology
PDF (1554 KB)


Automated Speech Recognition has many open problems. In this thesis two well-known problems are researched. The first topic deals with the ever growing phenomenon of English words being used in Dutch colloquial speech. This makes it increasingly expectable that speech recognizers in the Netherlands should be able to recognize these English words as well as Dutch. Dutch people speaking English introduces a number of problems, including non-native speech and lack of training data. Thus a multilingual acoustic modeling approach was attempted using training data from widely available Dutch and English corpora. The second topic deals with the realization of a call classifier which is trained using machine learning techniques. Machine learning techniques algorithms in call classification literature such as BoosTexter generally require a large amount of data before satisfying results are obtained, thus making them unsuitable for use in the beginning stages of a project where human expertise is more suitable. By using the RIPPER rule generating algorithm hopefully a system can be made which accepts human knowledge in the early phases of a project, as well as machine generated rules which can be inspected by the expert in later stages of a project as more data becomes available.

blue line
University logo