Speech Recognition using Neural Networks

title: Speech Recognition using Neural Networks
author: H. Turksema
published in: 1996
appeared as: Master of Science thesis
Delft University of Technology
also as: Alparon report nr. 96-04
Section of Knowledge Based Systems
Faculty of Technical Mathematics and Informatics
Delft University of Technology


"Speech recognition using Neural Networks" is a hype, which will probably prove its value in the near future. More and more researchers and companies are interested in computer applications which can evaluate and interpret spoken sentences.

The Knowledge Based System Group of the department of technical Mathematics and Computer Science at the TU-Delft understood the importance of this research area, and started almost a year ago, a research group called Babylon. Because of the lack of knowledge of Speech Recognisers the main idea was to reproduce a number of speech Recogniser which had already proven their value. In this thesis we studied a Neural Network Speech Recogniser as it was proposed by British Telecom (P.C. Woodland).

Before a Speech Recogniser can start its process of speech recognising a speech waveform has to be transposed into a number specific features. These features can be extracted from a speech waveform in many different ways. The most commonbly used methods are the methods which are based on the physiological speech producing and perception methods as we find them in the human body. With the help of the extracted features a recogniser can start to classify these features end decide which, for instance, letter was spoken. This classification can be done in various ways. Feature classification is most of the time done statistically.

Instead of difficult statistical calculations we tried to reach the same result with the help of a Neural Network which could 'learn' to extract these characteristics from a number spoken utterances. The Neural Network seems to reach a very good accuracy percentage. But because of the lack of good comparable results we have to be cautious to be too optimistic about using the Neural Network Method.

Finally it seems useful to look ahead and keep up with the developments in the research area of feature extraction. Feature extraction can be further improved with modern technics to obtain better results. This area is not extensively exploited yet but it shows some promising aspects for further developments.

blue line
University logo