Phoneme recognition using an auditory model and a recurrent self-organizing neural network

Neural networks that use unsupervised learning were used on the output of a neurophysiologically based model of the auditory periphery to perform phoneme recognition. Experiments which compared the performance of a recurrent self-organizing feature map to that of the standard Kohonen self-organizing feature map show that the recurrent version performs significantly better (t-test, p<0.05) in terms of phoneme recognition accuracy (30% vs. 25%) under the conditions tested (high signal-to-noise ratio and five sentences from each of ten speakers). However, the two representations make different types of broad class errors. The recurrent neural network classifier has smaller codebook size and lower entropy than its nonrecurrent counterpart