SUPPORT VECTOR REGRESSION FOR TONGUE POSITION INFERENCE
The articulatory inversion task consists on recovering the articulators’ position or the vocal tract shape from the acoustic speech signal. The availability of large corpora of parallel acoustic and articulatory data has made possible the use of data-driven methods as an alternative for the solution of the speech inversion problem. This paper presents a method for the inference of tongue positions based on support vector regression techniques. The acoustic speech signal is parametrized by using perceptual linear prediction coefficients (PLP); than, a nonlinear transformation function is applied to the regressors. Model assessment is performed by measuring the similarity between the estimated and the reference signals and by measuring the correlation between inputs and residuals. The proposed method shows to be promising.