PhD dissertation title:

Articulatory space modelling using a hypercube codebook for acoustic-to-articulatory inversion


Advisor : Yves Laprie


During my doctoral education, I conducted research about the articulatory-to-acoustic inversion (i.e., recovering the trajectories of articulatory parameters given an acoustic signal). Three steps were involved in this process:

(1) Modelling articulatory space by hypercubes,
(2) Retrieving all the solutions,
(3) Recovering articulatory trajectories varying slowly.

The inversion method I developed is based on the representation of the articulatory space by a hypercube codebook (Ouni & Laprie 99). This representation has the advantage of decomposing the articulatory space into regions where the mapping is quasi-linear. Each articulatory region is represented by a hypercube.

The inversion procedure retrieves articulatory vectors corresponding to an acoustic entry from the hypercube codebook. Since the dimension of the articulatory space is greater than the dimension of the acoustic space, the corresponding null space is sampled by linear programming to retrieve all the possible solutions.

Once the second step of retrieving all the solutions had been completed, these solutions could be used directly to study vowel articulatory variabilities by considering principal constriction place in the vocal tract to distinguish different classes of French vowels.

The final step, retrieving articulatory trajectories, was performed by using a non-linear smoothing method based on dynamic programming followed by smoothing with a variational method. The method used to retrieve articulatory trajectories was successful (i.e., smooth and realistic) as confirmed by a series of experimental evaluations.

Very soon, I will add in this page some examples of the inversion and some animations.