Speaker independent phoneme recognition with an auditory model and a neural network: a comparison with traditional techniques

Experiments were conducted that compared the phoneme recognition performances of two different preprocessing methods and classification schemes. Results showed that both representations (the auditory model and discrete Fourier transform) and classification schemes (self-organizing feature maps and K-means clustering) perform equivalently in terms of phoneme recognition accuracy (30%) under the conditions tested (high signal-to-noise ratio, average spectral vectors, and five sentences each from ten speakers). However, the two representations make different types of broad class errors. Both the auditory model representation and the neural network classifier have an advantage in providing codebooks with lower distortion and higher entropy than their counterparts