In this paper we describe a TDNN based hybrid word recognition system with a novel phoneme representation based on phonetic features. This new representation is more compact than the traditional 1-out-of-N phonetic representation and leads to a smaller network. In different experiments on a spelling letter database we show that the set of phonetic features has to be chosen carefully to achieve good results. We compare the new representation against the standard phoneme representation in the same experiment and show that the phonetic feature representation leads to better recognition results and more stable learning. With the phonetic feature representation we reached a word recognition rate on an independent test set of the spelled letter task of 96.1%.

Keywords: Word recognition, time delay neural networks, phonetic features, hybrid systems.