Acoustic-to-articulatory inversion

Abstract

Acoustic-to-articulatory inversion consists in recovering articulatory data, or the geometry of the vocal tract, from the audio recording of the speaker. The inverse problem may be formulated thus way: let $\mathbf{s}$ an acoustic vector containing the acoustic features observed in the acoustic speech signal (e.g. the formant frequencies), $\mathbf{p}$, the articulatory vector to be recovered, containing parameters that defines the geometry of the vocal tract (e.g. the area function), and $\mathcal{L}$ an operator that gives the acoustic vector as a function of the articulatory vector, hence
$$\mathbf{s}=\mathcal{L}(\mathbf{p}).$$
Then, acoustic-to-articulatory inversion consists in recovering $\mathbf{p}$ from the observation of $\mathbf{s}$.

In [1], I proposed a method to quickly estimate the area function and length function of the vocal tract of a speaker from the knowledge of the formant frequency embedded in the original speech signal uttered by the speaker. The method is an iterative method based on the sensitivity functions of the vocal tract [2] and weighted penalty terms for better regularization.

A few examples

simulation (/i/)

ArtSynthSchema

On real speaker: French vowels (/i/, /e/, /a/, /u/)

ArtSynthSchema


You can download the Matlab code here for acoustic-to-articulatory inversion of oral vowels.

In this archive, you will find a file test.m, which contains 2 examples. You may choose the inversion of a /a/ by choosing "load Library/a" and a /i/ by choosing "load Library/i".
In the Library folder, the constantterms.m file contains the constant terms used to compute the transfer function corresponding to the current area and length functions, using the chain matrix paradigm by Sondhi and Schroeter [2]. The default file contains the parameters defined by the authors. Feel free to change them. Updates are coming soon.

Please, do not hesitate to report any suggestion, dysfunctionnement, or weird result, to benjamin.elie(at)inria.fr.

[1] Elie B., and Laprie Y. "Audiovisual to area and length functions inversion of human vocal tract". EUSIPCO, Lisbon 2014.

[2] M. M. Sondhi and J. Schroeter, "A hybrid time-frequency domain articulatory speech synthesizer", IEEE Trans. Acoust. Speech Sig. Process. 35(7), 955-967 (1987)

Last modification: June 23, 2016