A Sign-to-Speech Translation System

This thesis describes sign-to-speech translation using neural networks. Sign language translation is an interesting but difficult problem for which neural network techniques seem promising because of their ability to adjust to the user\u27s hand movements, which is not possible to do by most other techniques. However, even using neural networks and artificial sign languages, the translation is hard, and the best-known system, that of Fels & Hinton (1993), is capable of translating only 66 root words and 203 words including their conjugations. This research improves their results to 790 root signs and 2718 words including their conjugations while preserving a high accuracy (i.e., over 93 %) in translation. The use of matcher neural networks (Revesz 1989, 1990) and asymmetric Hamming distances are the key sources of improvement. This research aims at providing a means of communication for deaf people. Adviser: Peter Z. Reves