|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Pathological speech processing
Participants: K. Daoudi, G. Li, Q. Robin, F. G. Satsou.
-
Small amount of training data in learning robust classifiers for differential diagnosis between progressive supranuclear palsy (PSP) and multiple system atrophy (MSA). We showed that factorial discriminant analysis and logistic regression can lead to such robust classifiers. Moreover, we showed that these models provide good insights on the multivariate variability and (un)correlation of acoustic features, which can facilitate clinical interpretation.
-
We investigated the problem of extracting ground thruth of glottal closure instants (GCI) from electroglottographic (EGG) signals of healthy and pathological speakers. We carried out a large experimental study which showed that existing methods are not robust to recording settings and material. We then proposed a method to overcome this problem. On the other hand, this problem highlighted the non robustness of state of the art methods in automatic detection of GCI from speech.
-
We made an experimental evaluation of state of the art methods in automatic extraction of the excitation source from voiced speech. To carry out this evaluation, we used a very recent source-filter model of sustained phonations. The results showed that these methods are reliable only in very particular cases and fail in most.
-
Matching pursuit (MP), particularly using the Gammatones dictionary, has become a popular tool in sparse representations of speech/audio signals. The classical MP algorithm does not however take into account psychoacoustical aspects of the auditory system. Recently two algorithms, called PAMP and PMP have been introduced in order to select only perceptually relevant atoms during MP decomposition. We compared the performance these two algorithms on few speech sentences. The results showed that PMP, which also has the strong advantage of including an implicit stop criterion, always outperforms PAMP as well as classical MP. We then raised the question of whether the Gammatones dictionary is the best choice when using PMP. We thus compared it to the popular Gabor and damped-Sinusoids dictionaries. The results showed that Gammatones always outperform damped-Sinusoids, and that Gabor yield better reconstruction quality but with higher atoms rate.