Artificial neural networks for phoneme recognition
P.T. Brunet, A.S. Pandya, C.V. Pinera · 2002
This paper describes the use of a backpropagation artificial neural network (ANN) to recognize sustained phonemes. The inputs to the neural network were taken from 74 points of an LPC spectrum. This LPC data was augmented by adding slope information to each point in an attempt to add knowledge of the shape of the spectrum. The approach was verified by merging the ANN into an existing speech therapy product, IBM SpeechViewer II, and then testing the ANN with a number of male and female speakers. Results are shown which demonstrate the viability of the approach. It was also discovered that the ANN was able to function in a speaker independent manner. However, results are also shown which point out limitations of ANNs in classifying phonemes which are quite similar such as the m and n phonemes.>