Isolated word recognition of deaf speech using artificial neural networks

Describes the development and implementation of a prototype speech recognition system for carrying out isolated word recognition of deaf speech. The recognition system is built around the TMS320C30 DSP processor using a combination of artificial neural networks and conventional signal processing techniques. A vocabulary of 50 words selected from the modified rhyme test produced by six profoundly deaf subjects was recorded. The acoustic, temporal and segmental characteristics of their speech were studied to identify features that may be useful in improving the performance of the speech recognition system. These features may account for the high variability and production errors in deaf speech. A recognition model was built using artificial neural networks to control the variability, and to use the information concerning the acoustic, temporal and segmental errors in deaf speech in the recognition strategy. The recorded speech samples were randomized and separated into varying numbers of training and testing sets. Selected features were extracted from speech samples and used to train the network to recognize the speech characteristics of each speaker, in a supervised learning fashion. The ability of the network to use the additional features to carry out speaker dependent speech recognition was evaluated.