Previous | 
Home
Section: New Results
Nonlinear Speech Analysis
Participants : Vahid Khanagha [correspondant], Khalid Daoudi, Safa Mrad, Nicolas Vinuesa, Blaise Bertrac.
-
MMF for speech analysis : we continued our research on the adaptation and application of the MMF to speech analysis and started a research theme on pathological voice analysis. We proposed a novel a compact representation of speech which consists in reconstructing a speech signal form its most singular manifold. This leads us to build a speech waveform coder which outperforms the G.726 standard. We then used our recently developed algorithm for Glottal Closure Instants (GCI) detection to improve the performance of our sparse linear prediction method. We also used this algorithm to develop new acoustic perturbation measures for normal/pathological voice classification.
-
Matching pursuit for speech analysis : we first showed that the Gabor dictionary is actually more efficient than the Gammatone dictionary for speech coding using the matching pursuit (MP) algorithm. This results mitigates some famous findings on the neural coding at the human auditory nerve. Second, we shoed that one single parameter, derived from MP decomposition of speech, allow discrimination between normal and dysphonic voices with an accuracy which is significantly higher than all existing methods.
Supporting grant: Inria CORDIS.
PhD thesis defended: Vahid Khanagha Novel Multiscale Methods for Nonlinear Speech Analysis , University Bordeaux-1, PhD defended on January 16th, 2013, supervisors: K. Daoudi and H. Yahia [13] .