Formant estimation from cepstral coefficients using a feedforward memoryless neural network

A method is described for estimating formants from cepstral coefficients using a memoryless feedforward neural network. The method was tested with vowel data from a large varied database. The neural network provided significantly better estimation of the formants than was possible with a linear transformation. However, the degradation in performance between training and test data suggests that the relationship between formants and cepstral coefficients is very complex and requires an extremely large amount of data for adequate neural network training.<
>