A digital neural network approach to speech recognition

Abstract

This thesis presents two novel methods for isolated word speech recognition based on sub-word components. A digital neural network is the fundamental processing strategy in both methods. The first design is based on the 'Separate Segmentation & Labelling' (SS&L) approach. The spectral data of the input utterance is first segmented into phoneme-like units which are then time normalised by linear time normalisation. The neural network labels the time-normalised phoneme-like segments 78.36% recognition accuracy is achieved for the phoneme-like unit. In the second design, no time normalisation is required. After segmentation, recognition is performed by classifying the data in a window as it is slid one frame at a time, from the start to the end of of each phoneme-like segment in the utterance. 73.97% recognition accuracy for the phoneme-like unit is achieved in this application. The parameters of the neural net have been optimised for maximum recognition performance. A segmentation strategy using the sum of the difference in filterbank channel energy over successive spectra produced 80.27% correct segmentation of isolated utterances into phoneme-like units. A linguistic processor based on that of Kashyap & Mittal [84] enables 93.11% and 93.49% word recognition accuracy to be achieved for the SS&L and 'Sliding Window' recognisers respectively. The linguistic processor has been redesigned to make it portable so that it can be easily applied to any phoneme based isolated word speech recogniser.