Player is loading...
ESR02 : Automatic detection and classification of pathological speech conditiond based on emotion expression
Zhao Ren
Thursday, Sept. 5, 2019 · 4:34 p.m. · 19m 27s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
either on or to repair work oh my word you're h. d.s daddy a
00:00:07
i am tower in uh e. i start in work high page one
00:00:13
my holstein state your case you the boys through a box perky in germany
00:00:17
uh my simple i and supervised by a perplexity on shore and doctor nicholas women's
00:00:25
next uh let me introduce my work from these or hard
00:00:30
first they've got up dat tapes of my work
00:00:36
uh my title uh well that that that's workdays automatic detection and that
00:00:42
a gentle right the logical each and asian it on emotion it right then
00:00:47
there are four i've created apes into great part
00:00:51
the partly that title page recalls beach a dagger may not with a chair
00:00:57
with a depression and bipolar let the if the if the if the if
00:01:02
and the second they've emotional expression uh emotion contents a continuous
00:01:09
like arousal and violence or uh that big parade
00:01:14
to emotions like hibernate and our fan ace and surprise
00:01:21
final way what is the refit relationship between pathological speech
00:01:27
and emotion expression i will use paper i mean apology
00:01:32
such as b. uh and our uh uh north in an uh to
00:01:37
investigate either into shape the uh the
00:01:41
relationship between emotion and a pathological speech
00:01:47
and go now uh uh in the first thing
00:01:50
here uh i have some publications and the networking activities
00:01:58
next i will introduce the research and publications first
00:02:04
and then later i will introduce than that working
00:02:12
uh in this one year a banana have player uh there are
00:02:17
taught to play one journal article and a six conference's papers haven't aged
00:02:23
today i will enjoy adios the final until a conference papers
00:02:29
to you
00:02:32
first they if the uh evaluation not hand level from speech
00:02:38
we cooperate way that oh yeah that was day of
00:02:41
will put how a tool clack to the pen data
00:02:46
and that the the uh a pen level evaluation
00:02:53
in our daily life um handy it's all
00:02:58
we can always be old pen and uh
00:03:01
a penny unhappy if if you are they
00:03:05
uh an an auxiliary role in that out
00:03:08
of my faith and i fighting death a the arteries of thirteen at all achieve such as
00:03:14
mental beefy this slight depression and uh um physiological
00:03:20
uh may be the if it's like kinds there
00:03:26
hi lara your day we use mine you know they're the most this like south reports
00:03:32
to uh evaluate the pen level but
00:03:36
it is time consuming and the costly
00:03:40
therefore in this work we tried to crack but they'd
00:03:44
car and uh use the automatic take evaluation might there'd
00:03:49
to make this work more objectivity
00:03:53
and high uh uh at q. accurate
00:04:02
basic now or a database
00:04:06
also it off at cute hancock has it tastes reported being
00:04:10
told whole three hours and the uh mm sample rate is um
00:04:18
one paid forty four point one a heart
00:04:24
oh we split the data into trendy that out and and high speed data on it
00:04:32
laura clancy it's waste meets the data in tool mailed moderate and it's the the uh
00:04:38
can i say it's depending on the uh um pen level from the our role to ten
00:04:46
so the the uh three cats classification task
00:04:52
two okay by a benchmark of these database we use
00:04:58
that's my third uh here all enough feature extraction part
00:05:05
there are two branches one is the uh i out p.
00:05:10
features we extract the company or a features and that's the the and
00:05:16
the spectrum features we then by date in tool bag of audio wired
00:05:22
too bad the uh it's a histograms of this features
00:05:27
that's that can run it on
00:05:31
we extract that come have featured and that's the thing features ah
00:05:36
and need actual graham uh the features i thought a segment reliable
00:05:44
and then follow the had occasion we pay pay its features into x. p. m. or
00:05:50
i. r. s. t. m. for that by no path haitian for the experiment no settings
00:05:56
uh oh we sat there as they
00:05:58
um a complexity parameters the burial in ten
00:06:03
a bit and an how of minor sticks
00:06:07
and a ten mile how well minus one
00:06:10
uh with that they're asking m. r. and i'd work as the railway ours
00:06:18
in the experiments we can see that
00:06:23
uh this
00:06:25
compare features as i say i am and a bag of audio words
00:06:29
that's that ice the um i have worked there well at a hot
00:06:34
when we use as they am and uh they expect from
00:06:39
features work well it well we use our asked he um
00:06:44
and finally uh the thief that in your own magic's we can say that
00:06:49
the um most of these are are it's classified as the first the
00:06:54
class may older we think this leave because of this data is not balanced
00:07:01
the mayo the data is more than moderate and they say they are the top
00:07:09
so this this is mark we gave up benchmark up and data
00:07:13
next though let me introduce the uh seconds daddy
00:07:19
marty is is going for a bipolar disorder diagnoses
00:07:27
uh your native learning uh as we know people earning kind on my a. b. worked me uh
00:07:34
my name's large databases although eighty eight height already got great for
00:07:40
five in miami beach it the
00:07:42
learning have like speech recognition speech generation
00:07:48
however in pathological speech we always
00:07:52
of place my knee problems that the data is that if our with more
00:07:59
not only the individual number of individuals are limited
00:08:05
but also hold a total they have number are not aware of it
00:08:12
therefore which might to find a a an approach
00:08:17
to solve this problem use the learning along some more data set
00:08:28
a bipolar disorder corpus here we use a is uh i
00:08:34
didn't have that high pitched in at that challenge last the year
00:08:38
it tastes a bit has that require date
00:08:42
wrong forte fix turkish speaking in the beatles
00:08:48
and the the up that hole are they saw their level was evaluate
00:08:54
hated by the your money or rating scale then in this had arranged this
00:09:01
why i am our eyes scale was split it
00:09:05
into three classes remission happen mine yeah and money on
00:09:12
with with that it had this to the ocean
00:09:17
uh as weak as the uh the data is uh almost to a balanced here um
00:09:25
oh the ideal here uh in the training that twenty by thirty eight
00:09:32
forty line uh these values are the
00:09:35
original speech uh clips and then in the
00:09:41
you and the bigger numbers are the uh speech trunks uh out here in
00:09:47
this table they are the us a number of friends of these speech create
00:09:58
in our uh remark market instance learning framework
00:10:05
we think the market is the it's learning a it's a meeting to solve
00:10:10
the problem of weekly label which lead
00:10:14
ha with clay labelled data is the um
00:10:19
here in this problem means that the audio uh up waves are
00:10:25
always if a very long uh near earning more than ten minutes
00:10:32
um it was not very well annotated so weeping each the only way
00:10:38
if i thought map and there's that meant eight into that we're always here
00:10:46
and i think that the small errors all the all segments as instances
00:10:54
for their uh from these audience that means we can't that i must face the features and uh
00:11:01
using the instance level classification remark to paper prediction
00:11:06
to each and that's the cities around the audio segment
00:11:14
next so that's the final prediction
00:11:18
we use that level of classification tool fuels all of the youth predictions
00:11:28
that level kite cajun and tens of your assumptions
00:11:32
we all had all of these or assumptions that rates
00:11:36
maybe eighteen now we're uh experiment tool gave and of comparison
00:11:43
its contents and a standard assumption uh which means that uh if
00:11:50
if uh this audio wave original audio wave ace predicted as eight
00:11:59
that's means there is there are actually is the
00:12:03
one prediction of from the all the old segments
00:12:10
ace pretty dictate as a cell here we use the max
00:12:15
rule to gape the final prediction and the set an ace
00:12:21
vocabulary based on that assumption here we use a histogram tools that
00:12:27
his big the predictions and then choose the maximum number in the histogram
00:12:35
and the third they've cracked table of function this means that
00:12:39
uh all the piece that man's hand tape the same contribution
00:12:43
to the final prediction so here we use the mean you
00:12:49
finding a palpable back level kath
00:12:53
patient also quentin way taped active assumption
00:12:58
it means a on the beef audio segments are ham paperless uh
00:13:06
can come to builds difference to the panel prediction
00:13:11
so we knew the paper out weighted down of all of
00:13:16
the probabilities of your predictions here to compute the final prediction
00:13:24
in the experiment to
00:13:27
we also will improve the marquis instant learning using example learning
00:13:33
we opted we can't attend aside applied to
00:13:36
pay hours from the two different team chaney iterations
00:13:41
and then we use the high level that haitian might third too bad all up these out will
00:13:49
to tape the final prediction for that we'd had gave an improvement
00:14:01
monday's table we can see that the example
00:14:04
marketing and learning no my terribly e. which
00:14:10
uh back level classification my third week use example marty he's this learning and
00:14:16
they an improvement of marty expense learning
00:14:22
for their week on higher our my third with either state of the art might hurt
00:14:28
i'm also working on a bipolar disorder of
00:14:32
hopeless your rain i backed challenge lights here
00:14:38
uh in the all of the my third skews seeing over they
00:14:43
all deal here i were my thirtieth higher than bills might hurt
00:14:48
and the of also all our martinis is learning my third hand
00:14:55
they if the same pro performance with a trend here um
00:15:01
uh uh have a remote nicole week or models using old audio and video data
00:15:09
um that had that that
00:15:14
all this food that's that and id
00:15:17
and then let me introduce the night working
00:15:23
um in the past when you're a a joint at the high
00:15:28
pass twenty events line and to and to re in this two days
00:15:33
i was all last year i join tightly g. conference and
00:15:38
the case workshop uh in this year i do and icassp conference
00:15:45
tool gave 'em presentation and i also joins summer school ah
00:15:51
in july of this year i also open mine interesting topics
00:15:58
your rain this done with such a it's a it's expandable a. i. and
00:16:04
oh come for a son of a cop comprise an uh a neural networks
00:16:13
and uh uh also in the ah activities we the white said
00:16:18
to us a constructed by a a you really think you were much
00:16:23
two parents to provide as a night form to prevent our work
00:16:31
and uh in you the unabomber say a
00:16:35
box art work we uh i you know
00:16:41
oh right to to eating quality of writing mobile application development and
00:16:48
also the two tower will of speech pathology and the practical um
00:16:55
a mobile fencing for fitness and health and not
00:17:00
to tell real well speech pathology uh was ah
00:17:05
the latter was gave them by doctor meek and
00:17:09
uh i gave a tutorial for the whole semester
00:17:17
in the future
00:17:19
uh oh
00:17:21
i will call we will cooperate with e. d. out and can be normal university tool
00:17:27
uh released a new database about it actively
00:17:31
than a high that up the today and
00:17:37
which which is also a very interesting topic
00:17:42
tool mm impact the gate to evaluate though
00:17:48
can i have to leave kind the kind that's liable to how that doctors to
00:17:53
pave um of older a tag another
00:17:56
thieves and go off operation decision um hum
00:18:04
i mean the might there it's part i will
00:18:08
trying to use strands were learning tool
00:18:13
saw the it's more dave how problem and the
00:18:19
that can dates expandable a i i will use eight
00:18:23
such as uh attention mice mansion needs them to all and the
00:18:28
black box of ignore networks to have a little what we have learned
00:18:34
from the deep your networks for speech pathology a ravenous it's
00:18:44
and uh i am of conference faith and journalists
00:18:47
around include a includes at cats enjoy speech and the
00:18:53
transactions on multimedia these are my future lines to that made
00:19:02
and uh in the uh
00:19:04
oh to them and winter of next year i will go to all to remain
00:19:10
two two mice a calm and
00:19:13
pondered a supervisory seeing of grief for
00:19:19
uh_huh
00:19:22
thank you for your time and attention had time
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.
161 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
End-to-end approach for recognizing speakers from audio
Subhadeep Dey, Idiap Research Institute
April 19, 2018 · 11:09 a.m.
232 views