We showed a new scheme to characterize speech from LSP parameters by 5 layers sandglass type nonlinear neural network (SNN(NL5)). In order to synthesize speech, we take advantage of useful abilities of SNN(NL5) for compressing and restoring the information. We performed learning experiments on LSP parameters of 5 vowels to investigate the ability of SNN. The followings were verified, 1) the distribution of LSP parameters compressed by SNN(NL5) are similar to the distribution of F1-F2 formants plane. 2) Nonlinear output function of neural elements in second and fourth layers of SNN(NL5) work effectively from view point of separating the distribution of vowels. 3) In order to prevent SNN(NL5) from over learning, there exists the optimum numbers of neural elements in second and fourth layers. For 14 orders of LSP parameters, this number was determined to be 20. 4) There is a preferable property on the plane to separate the vowels distinctively when the restoring error of LSP parameters becomes less. 5) SNN(NL5) can restore the LSP parameters with accuracy enough to synthesize speech from the compressed parameters.
