I am interested in articulatory modeling and speech analysis with a view of better understanding how speech is produced and of identifying the acoustic cues supporting the phonetic contrasts.
Articulatory synthesis
Currently I am working in the field of articulatory synthesis with of view of synthesizing realistic
speech signals from the knowledge of a sequence of phonemes. There are two main stages :
- 1) generating a geometrical shape of the vocal tract which changes over time. There are roughly
two time scales that of speech sounds one every 80 ms on average and that of very brief acoustic events
as a transient noise when a constriction is released.
- 2) generating a sound by simulating the vibration of the vocal folds and the propagation and
modification of this wave through the vocal tract.
For the first aspect we exploited old X-ray films from IPS
(
Institut de Phonétique de Strasbourg) and I developed Xarticul software to extract the contours of each articulator (tongue, mandible, lips, larynx, velum and palate). These contours were used to design an articulatory model which describes the shape of the vocal tract from a small number of deformation modes. This models has to be controlled from the knowledge of sounds. This requires the prediction of articulatory gestures which enable the production of the sounds with the expected acoustic properties.
For the second aspect we are using acoustic simulations realized by
Benjamie Elie.
This work was initiated within the framework of a
LORIA project
using sequences of VCV (Vowel Consonant Vowel) and is now carried out in the framework of the ANR project ArtSpeech.
Acquisition of articulatory data
This work is not imaginable without articulatory data and we are thus collaborating
with
LPP in Paris to realize aerodynamic
measures and capture vocal fold parameters and with
IADI to acquire MRI images and cineMRI data.
Speech analysis and language learning
In the domain of acoustic cues I previously design algorithms to extract F0, formants and burst characteristics
from speech signal. These algorithms were implemented in the WinSnoori software. We are currently using
this knowledge in the domain of language learning (
IFCASL project) to provide learners with acoustic
feedback which is intended to guide them toward the realization of the expected sound.