Player is loading...
ESR15 : Automatic assessment of speech in cochlear implant (CI) users
Tomás Arias-Vergara
Thursday, Sept. 5, 2019 · 5:32 p.m. · 14m 15s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
hello everyone um and then use the messiah yes i'm i'm there yes our uh fifteen
00:00:07
uh i assert a than a deposit let
00:00:10
your mail as your um michael's institution with a
00:00:13
little bit months in union a university if particular area on uh the clinic of the university
00:00:20
so um i supervise or star my yeah and my um um bravo
00:00:26
so the title or the topic of my uh prodded
00:00:29
is the automatic assessment of the speech of poplar implant users
00:00:34
so the main ah to updated of my prodded this to use
00:00:38
of speech at the technology to help but it's not a fifth
00:00:43
also tool incorporate this uh speech data into the clean if the the patient
00:00:49
theft so i'm going to give you a short introduction uh just
00:00:54
if you don't know all that concrete plans are applying medical devices star you'll still
00:00:59
provide some data fields protection if two
00:01:02
people that hostage here or profound hearing loss
00:01:06
so it's basically a only a process or with a microphone which a
00:01:11
combat they our real signal to electrical signal that is then ah does need
00:01:18
electrode um and then is stimulates the the the nerves that are in charge
00:01:24
of send in the acoustic or auditory information to the right to the brain
00:01:31
ah where where you went after they're complying plantations on out the patient my
00:01:36
ah save or not is some differences between the sound preview and the sounds operative
00:01:42
so it makes sense because for you some when you are lonely in a a new language when you hear
00:01:48
the correct pronunciation of or you say it and you try to to to say yes similar out that's you hear it
00:01:56
at the current their pronunciation but you you heart problem
00:02:00
with your your hearing you might up reviews are different found
00:02:06
so um small so the speech problems um uh on the cochlear implants
00:02:12
our uh a decrease in thirty intelligibility
00:02:15
ah changes in terms of articulation ah so
00:02:19
not then has to increase or decrease of nasality and also is lower speaking rate
00:02:25
and b. is uh the type of the problem
00:02:28
the speech program on the disability might depend also
00:02:32
on whether grip patient on lot they viewing before or after uh the below may know it's speech
00:02:38
so if you well star before so the burning well uh on sub pop of the night
00:02:45
uh people actually never go to to learn how to produce the the sound correctly
00:02:51
so they might have problems with with this and
00:02:54
the few the cave up uh after development of speech
00:02:58
so the the page can yes uh they learn how to play with the sound bad
00:03:03
and the labels like like the auditory feedback
00:03:08
like their training on how the sound is supposed to be produced
00:03:13
so the aim out the prodded it's to implement different methods
00:03:18
send a signal um signal processing and much alert mean to
00:03:22
about the way the the speech of the the cochlear implant users
00:03:27
so i'm going to give you are really and all that an
00:03:31
overview of what the methods i use it uh use it so far
00:03:35
so the fifth one if he if uh you see in
00:03:38
the the onset and all the transitions to evaluate the articulation problems
00:03:44
so basically an aunt a transition is the transition from now both latest voice sounds
00:03:50
and the outset these they boarded the transition from our voice to oppose let's um
00:03:55
so to get different patients um i compute
00:03:59
the fundamental frequency over the whole uh recording
00:04:03
then i look into the boundary at the deli meat or the
00:04:07
point where the a fundamental frequency it start in the it starts
00:04:11
uh up for the case up they on the transition so i get the the first point and
00:04:17
at this a boundary i i take out window above eighty milliseconds will lead on to the right
00:04:23
and then uh i start us on acoustic features from these translations
00:04:28
um perform some classification between cochlear implants on narcotic controls ah
00:04:34
to to know whether the onset or author translations are
00:04:39
just we're about to to the to the articulation problems
00:04:43
so the results from this uh first study were
00:04:46
already presented at that conference and partly let's make
00:04:50
and the best it's all where are playing with the onset
00:04:53
transitions so uh the core id or the the the main um
00:04:59
a conclusion used uh uh yes we can do that the articulation problems
00:05:04
r. and the onset transitions where the basic ones to to do this to perform this task
00:05:11
where this is very general we don't actually know um where the pages are making that mistake
00:05:17
ah regarding the articulation or the the production of of sort fine for me
00:05:23
so all the next uh instead was to perform some forced alignment
00:05:29
and label the the phonemes according to five by classes so
00:05:35
the travels where are an angel found phoebe land fricatives um well if somebody's laces stops
00:05:41
then uh i take the consonant well well um what was the
00:05:45
consummate transitions are you extract some for acoustic featured from the sensations
00:05:51
and again perform some classification uh between walker implant some kind
00:05:56
to have the controls in order to know um ah which
00:06:00
um uh phonological phonological club not now phoneme class is the one that that the
00:06:06
patients make weren't sake so i'm from the
00:06:11
resold we we saw that the the patients
00:06:14
or the the best with occasional fall where
00:06:17
ah yeah considering this stimulants to what what's transitions
00:06:23
and and the resource from this study are going to
00:06:27
be present on this week at the the is the conference
00:06:31
so now we know that we can do that the um the articulation problems and we also
00:06:37
no all that we can look out which show phoneme classes
00:06:42
are the ones that they're a patient makes out the more mistakes
00:06:47
and then the next uh a step was to look into features that will phone
00:06:53
me or give me an index or of how good or bad was the pronunciation
00:06:59
for these um i consider the political features
00:07:03
we'll re hear about this site yesterday but basically
00:07:06
what we went to to to do is um or
00:07:10
would what without without the net right now on their network
00:07:14
which is trying to get on a different ah phonological um features
00:07:19
then we get out sequence of acoustic features that
00:07:22
are compared to us sequence or on the local posterior
00:07:26
the sports posterior come to use as an index of how good or bad uh our sound is produced
00:07:32
so for you sent it deal looking to the to
00:07:35
the nato uh uh posterior of for the green uh ah
00:07:42
and lines ah we cannot or about that it's like are really straight goals
00:07:48
uh which means that the posterior for beef and
00:07:52
nasal sound is star one or cloth one so
00:07:57
we can say that they produce the production of these uh sound it's
00:08:02
laugh but also the if like ah
00:08:07
ah a nowhere ah i've fulfilled probability
00:08:12
so these my uh indicate that the sound was and stuff their faith i mean he might be
00:08:20
might be a point fire point save it doesn't matter
00:08:23
ah but it's not there for it and then we can say we can take these
00:08:27
are are some measure of us an index of how widow or how about was produce
00:08:34
and um especially because when we analysts um up a speech of poker implants
00:08:41
we can get or might get a posterior it will work
00:08:45
on this if we compare for instance if this is it
00:08:49
so the the souls for for these studies are one percent um the future speech
00:08:57
and in in two weeks um
00:09:02
another uh study or or experiments on perform what's um also on
00:09:08
yeah the blog mean a base approach it well it's i convolutional
00:09:14
neural network with a two channel or spectrogram so the
00:09:17
input classified between cochlear implants on our of the controls
00:09:23
so and the idea i use uh we have they the the input
00:09:27
which are two channel is spectrograms in one channel we have found mailer spectrum
00:09:32
and the second channel we have got that much on a spectrogram
00:09:36
so the difference ease of analysis actually uses oh triangle artist or band
00:09:41
and how much on ah the shape of the filter bank for
00:09:45
the tomato not feels them of functions and mortified five a sine function
00:09:50
and um the the this study is also
00:09:56
what was needed for for a conference in butter
00:09:59
recognition this is more a general study and and i will be presented this in the tower
00:10:07
so for my current work um now i'm working on 'em automatic
00:10:12
method to to the third methodically problems says that it is he
00:10:20
so the idea is i have ah fifty
00:10:27
not particular about so i have a normal synthetic speech as a baseline
00:10:33
ah i have the s. d. synthetic speech uh some late in um
00:10:38
no silently label for when the bell when you stop in it or press and then fight then seven and ten
00:10:47
and the idea is to take a these ah the recordings these
00:10:51
synthetic speech as a baseline being ah obtains on the speaker models
00:10:57
for three different groups so we have a we have the hope
00:11:00
the controls really well contrary one sample inward popular in plain view source
00:11:06
and the and ah we come and compute the baby ifs ands or the
00:11:10
difference between the baseline model and a speaker more work so we have normal
00:11:18
top
00:11:28
other and do the same for the by seven and and or ten percent
00:11:35
ah so the if uh actually up really preliminary resold soul
00:11:41
we have um five boxes for for uh the the tree a
00:11:46
a speaker groups or how to control a printing was compelling while
00:11:50
the distribution of the uh a good boxes are the different between the speaker models
00:11:57
ah respect to the norm our a synthetic speech
00:12:00
then digital one thought that the fan between the speaker
00:12:03
model respect to an awfully gullible low throughput phone and
00:12:07
so on and we have this for for every group
00:12:11
so it's very in three e. three if very uh nice to see that belief kind of a pattern that
00:12:20
that makes sense because twenty five for big probably uh for the how to control
00:12:24
with the that they'd be fun for both higher higher ah depending on the not holiday level
00:12:30
so they hired the the d. stand in the
00:12:33
the more different the two ah am models are
00:12:39
so we can see that for they had to control the bagel
00:12:43
higher for the reading was it'd be different if closer when they
00:12:48
not how you can label two percent of computer uh for for the bill baseline
00:12:53
and also we cannot for the kind of thing
00:12:56
behaviour for the public what's all this might indicate that
00:13:00
ah they might have been a study to problems
00:13:02
but outside say um this is baby preliminary i
00:13:07
at the moment we are i'm collecting not collecting bills the label in
00:13:12
the data out to to get some clinic so the the value to able to evaluate these these approach
00:13:20
and also um i'm working on a method of for the boy
00:13:26
since i know onset time detection you soon recorder networks so um
00:13:34
yeah basically is
00:13:36
using nah as some
00:13:52
yeah yeah yeah it's good to be able to to
00:14:01
to measure articulation problems seen the patients out of well
00:14:05
and and yeah that's it so
00:14:10
you have any questions
Conference Program
Storage and management of sensitive speech data
Henk van den Heuvel
Sept. 4, 2019 · 11:51 a.m.
ESR01 : Speech based automatic dementia detection in a home environment
Yilin Pan
Sept. 4, 2019 · 2:14 p.m.
ESR03 : Interpretable speech pathology detection
Julian Fritsch
Sept. 4, 2019 · 2:30 p.m.
ESR04 : Acoustic speech based monitoring for telehealth and predictive analytics
V. Srikanth Nallanthighal
Sept. 4, 2019 · 2:54 p.m.
160 views
ESR07 : Modeling progression of patients with neurodegenerative disorders
Camilio Vásquez
Sept. 4, 2019 · 4:05 p.m.
ESR08 : Developing measurment procedure of pathological speech intelligibility
Wei Xue
Sept. 4, 2019 · 4:31 p.m.
ESR09 : Clinical relevance of intelligibility mesures
Pommée Timothy
Sept. 4, 2019 · 4:49 p.m.
Big Data with Health Data
Sébastien Déjean
Sept. 5, 2019 · 9:20 a.m.
Small Data, Number of features vs. number of observations
Olivier Crouzet
Sept. 5, 2019 · 9:47 a.m.
ESR11 : First year review
Bence Halpern
Sept. 5, 2019 · 11:20 a.m.
ESR10 : Acoustic features to support the perceptual evaluation of accent production in dysarthric speech
Viviana Mendoza Ramos
Sept. 5, 2019 · 11:39 a.m.
110 views
ESR12 : Continuous speech recognition for people with dysarthria
Zhengjun Yue
Sept. 5, 2019 · 11:50 a.m.
153 views
Management of health data in an industrial company
V. Srikanth Nallanthighal
Sept. 5, 2019 · 1:57 p.m.
146 views
ESR02 : Automatic detection and classification of pathological speech conditiond based on emotion expression
Zhao Ren
Sept. 5, 2019 · 4:34 p.m.
ESR13 : Bridging the gap between typical and pathological speech recognition
Enno Hermann
Sept. 5, 2019 · 4:56 p.m.
ESR14 : Developing speech therapy games for children with speech disorder
Thomas Rolland
Sept. 5, 2019 · 5:16 p.m.
ESR15 : Automatic assessment of speech in cochlear implant (CI) users
Tomás Arias-Vergara
Sept. 5, 2019 · 5:32 p.m.
Recommended talks
Assessment in speech disorders
Virginie Woisard, Toulouse
Sept. 25, 2018 · 2:45 p.m.
140 views