The difficulties of nasal vowel recognition are twofold. The additional zeros as well as poles in the acoustical speech, introduced by the connection of the nasal branch to the oral branch, make it impossible to acquire a stable and clear spectrum. On the other hand, analytical model are insufficient for explaining the zeros and poles appearing in the spectum. In this article, we shall discuss the general characteristics of french nasal vowels, and then introduce a pole-zero model in the frequency region selected by local optimization of the spectral matching. Lastly we combine the LP all-pole model with this local optimised zero-pole model for the recognition of a data base (325 words) of natural french nasal vowels prononced by thirteen male speakers.

In the experiment, we have not chosen any reference, because we want to find speaker independent parameters. The result of the recognition shows that the combination of LP model with pole-zero model is able to help us to reveal the parametric changes of the nasal vowel /5/ and /a/. As to the /e/ /oe/, we were unable to find more information about nasality.