[Applause]
[Music]
or it is basically just a hole by their
boot events so what’s happening right
now is I’m manipulating the shape of my
multi-track and that’s manipulating
physics and having a whole bunch of
acoustic events and you’re interpreting
that at speech so one of the things that
you could say happens in reception which
is just you’re suing acoustic pattern
matching you’re taking the frequencies
that are coming out of my mouth your
magic
you’re saying I understand what it is
that she’s saying and there’s a lot of
evidence that that is what humans do we
have auditory canal we have a bow chain
because amplifies frequencies coming out
of my mouth we have a cochlea that has
hair cells that librarians sympathy to
those frequencies that it’s connected to
some auditory nerves that tells our
brain which frequencies are being
created and magically somehow if you
hear a 300 Hertz and a 21-13 frequencies
that you say she was produce right so we
could say we’re just matching acoustics
and there’s no reason like Siri can’t do
that there’s no reason why Alexa can do
that speech recognition should be a
breeze well it turns out there’s a lot
more that goes into it and that’s what
I’ll try to convince you of today so
we’ll start with this quick little
experiment don’t let the sky play again
and what is it what is it that you’re
hearing
ah alright stop for one sec alright
we’re gonna play it one more time but
what I want you to do this time just
close your eyes all right close your
eyes and here we go again okay you can
open your eyes closed doing acoustic
pattern matching you’re taking the
acoustics and you’re brainless saying
match those to a pattern
I hear ba but when your eyes were open
what was happening was you are also
taking visual cues in in a nutshell this
is called the McGurk effect and in a
nutshell this is just your visual cortex
messing with your audio cortex so what
I’ll do for the rest of the day is try
over the rest of my luck not the rest of
the day you only have to say is to try
to convince you that when the whole
cerebral cortex gets involved it’s a
mess it’s a big mess but it’s a useful
mess all right so let’s go on to this
next thing this is it easy what’s the
past tense of how do you make a verb
past tense in English
Edie of course Edie what about spoken
English what do you add yeah duck right
hug
past tense hug what do you add to kiss
are you adding a D hearing a T right
what are your adding to Wade you’re any
good so you have two all right so it’s
not Edie it’s actually at least three
it’s actually a couple more but we’re
gonna stick with these three right now
so maybe what’s happening is you
memorize
in the English language and you memorize
every single past tense that goes with
those particular verbs but if I give you
word you’ve never heard before each one
of you is going to say the exact same
thing about what the past tense of
English is so the past tense of myth is
mid mid sorry mid a sense of meet with
the tea past tense of duck does it
alright you know do this alright so this
is something that you know about
language that you didn’t even know that
you knew and let’s do you think this is
a little weird thing with the past tense
same exact thing happens with the plural
you all know the plural of English in
English is to make a noun plural you
just add a s what about what do you add
to a cat s cats sure what about dogs
that’s a Z and in fact most of the time
we can make a noun plural you add a Z
right so you think it’s an ass but it’s
actually a Z and then to fish I don’t
want to hear fish is the plural of
fishes fish
and then you add right alright so yeah I
mean those right and then if I give you
words that you’ve never heard before
you’re all going to do the exact same
thing all right so the personal thing
that’s happening here is in our
conscious mind we’re sort of
oversimplifying things we’re saying at
several that we know that when we have
to add an alveolar to verb that ends a
voiced sound a daddy voices at a t if
it’s another IV later
know just make it good all right say
there’s a really set of complex rules
that tells us the plural no let’s just
simplify it and say that it’s s so one
of the things that our brain does is
takes this really complex system and
simplifies it what do we do with the
speech recognition system well we find a
linguist and we say help us with these
rules that are in people’s heads that
they don’t know that are there because
we’re going to need them to make our
automated speech recognition system work
well all right let’s do another little
test here I’m going to just say a phrase
what’s the phrase that I’m saying the
last speech so the last speech well
we’re at a TED talk so what you probably
heard me say was the last speech all
right but if we were driving down the
coast and I said this is the last speech
you wouldn’t hear it as speech you would
hear in his speech acoustically the last
speech and the last speech are exactly
the same they’re articulate orally the
same so their acoustic Li the same and
what you do is you use the context to
help you figure out whether you are
Beach
are you heard speech all right we do
this all the time can we build this into
a speech recognition system sure there
are probability trees or decision trees
that people are trying to build in the
system saying if these set of words
recognize then it’s more likely that
this word is this and not this so it’s
harder it’s more complex but we can
build that into the system all right
so you guys don’t have a problem with
this because you understand that context
is going to mean that I’m likely to say
the next speech if we’re here at a TED
talk and unless I’ve been nagging you
the whole way down our lovely Drive of
the coast you’re going to hear it as the
last beach and at the last speech all
right what else goes into this all right
in the 1950s people started using kind
of old-timey speech synthesis to do
experiments on the human perceptual
system one of the things that they did
was they realized that they could change
the precept of what vowel somebody was
hearing by manipulating a certain set of
things so one of the things that Varro’s
what vowels are basically just kind of
categorical boundaries in this big blob
of acoustic space so yeah that’s just
all a blob of a bunch of different
acoustics based on almost an infinite
amount of different acoustic patterns
are showing up but your brain says
divided the acoustic pie a certain way
but the frequencies that come out of my
vocal tract are going to be very
different from the frequencies that come
out of somebody else’s vocal tracts it’s
completely dependent on the size of your
mouth the size of your larynx how long
your vocal tract is what kind of what
your teeth are like all these kinds of
things affect the frequencies so somehow
we have to figure out a way as a human
perceiver to make sense of this and one
of the things it turns out that people
do is they pay attention to what the
range of frequencies are that are coming
out of their interlocutors mouth so they
synthesized with their old timey
synthesizer the word bet all right if
you have a vowel that sort of midway
between bit and bet you want to see what
people are going to perceive and it
turns out if you put this word in the
end of a sentence that has the top
frequencies people are going to say I
hear bet but if instead you have the
bottom set of frequencies people are
gonna say I hear bit in other words what
people do is they calibrate their
interlocutors vocal space they say
here’s the frequencies that I hear and
I’m going to contextualize all the
vowels that they come out based on what
I have calibrated for them there’s been
lots of studies since the 50s with much
better since speech synthesis showing
exactly this so one of the things you do
is you just say I hear somebody and now
I’m going to do some calculations and
use that information for how I interpret
everything else that they say we can
build this into a synthesizer – it’s a
little trickier but we can just have the
cyntha I means I’m sorry they a
recognizer we can have the recognizer
get the range of the pitch but somebody
is producing and we can work that into
our calculations do we do this at
another level sure one of the things
that happens is for most people males
have longer vocal tracks and longer
longer vocal cords than females what
they have found in a lot of different
studies is that just seeing a face that
looks female that people have identified
as female or people have identified as
male changes the precept of certain
things to give you one example the
difference between sip and ship so the
difference between and Shh
it’s just where that band of white noise
is first it’s in higher frequencies
first it’s in lower frequencies if I
synthesize a token that’s halfway
between sip and ship I can try to figure
out what it is people are going to do
with that and one of the things that
happens is if they think they’re
listening to a female they say oh that’s
ship
if they think the very same signal is of
coming out of a male mouth they say oh
that’s sip in other words even before
people have heard the frequencies coming
out of
their interlocutors mouths they’ve made
some decisions about what kind of things
will come out of that person’s mouth in
order in other words they sort of pre
calibrate and they do this with kids
because kids are gonna have even smaller
vocal tracts smaller vocal cords higher
frequencies they tend to look at
people’s height they tend to look at all
kinds of different things about their
interlocutor before they even open their
mouth all right can we build this into a
synthesizer sure I’m sorry I keep saying
synthesizer recognizer speech
recognition system can we build this in
yes but the ramifications of this have
been that for the first several decades
of work on automated speech recognition
male voices were recognized much better
than female voices leading one of my
colleagues to say in the early 90s I
don’t know what it is about women they
just can’t seem to be recognized by our
machines so in any case we fixed some of
this now so in the decades since it
started we fixed some of this alright
but here’s where it gets even more
tricky if you’re building an automated
speech recognition system we interact
with people with all kinds of different
dialects so to give you an example in
Houston some people have marched it and
ad in front of nasals in other words pin
and pin are homophones you stick with a
pin you write with a pen all right other
people have that as two separate phone
to separate lexical items pin pen all
right
who says pin pin who says pin pen well
one of the things we know is that for
Anglos older people are more likely to
say pin pin younger people are more
likely to say pin pen do people know
this well if you do a perceptual test
and give you something like rinse and
rinse will also be homonyms if you have
this merger they won’t if you don’t so
you’re either rinse and rinse or rinse
and rinse if I do a perceptual test
ask somebody to choose what they’re
hearing but I hooked them up to an eye
tracker one of the things I might be
interested in is where their eyes are
going well if I put the picture of the
older lady in the middle and say this is
who you’re listening to it turns out
that people’s eyes sort of move back and
forth between rinse and rinse if on the
other hand I put the younger picture in
there their eyes go right to prints in
other words you might not have known
this at a conscious level but people
react as if this information is in their
heads I have an idea that older people
are likely older white people are more
likely to emerge this and younger white
people are more likely to not so I’m
going to use that information all right
what do we do with this with this with
an recognizer
I’ll talk about that in just a second
but let me give you one more example
what else can we use well there is a
feature of almost every dialect of
American English that says instead of
saying alveolar stop at the end of the
word I can replace that with a glottal
stop so instead of saying what I can say
what so I didn’t break my time up to my
alveolar Ridge in the second one and the
first one I went what and bring brought
my tongue to the my alveolar Ridge in
the second one what I just closed my
glottis but you probably all interpreted
the second one as what so I guess all of
the people of American English got
together and said what how about if we
just say that the grotto can stand in
for the alveolar if it’s a teeth okay
[Music]
and there are some African American
that stayed back and said you know what
D is also an alveolar so how about if we
make not only the T stand in for the
alveolar offer they got up with the
glottal stop but also D so good is
interpreted as good but so is good if I
stop things at the alveolar Ridge I mean
at my glottis that can still stand in
for a D let’s do that though just taking
this same rule just pushing it one step
farther okay that sounds good should we
tell the white people for the Anglos
that can only be SCT for the
african-americans that could be SCT or
sa ID if we put the dude on the right in
the middle and say this is what you’re
listening to people pick set if we have
the dude on the left and say this is who
you’re listening to people pick either
said or set alright so this is probably
a rule that you’re not aware you if I
ask you what people do with alveolar
stops if I ask you what the difference
is between African American and European
European American English you might not
say this but it turns out people know
this at some level so what do we do with
our speech recognition system we kind of
say this is there’s no way we can really
build this into the model except to say
some people might do this and some
people might do this all right
so perceived age of the speaker
influences how we perceive things
pristine ethnicity so does gender lines
on a map and perceive disability and
those
three last ones are in red because we’ve
done some experiments showing that in
fact those things can even override the
acoustic signal I did a lot of work with
my people I grew up with in Michigan
because one of the things that I found
interesting was a lot of Canadian things
that people were associating with Canada
were creeping into anger speech in
Michigan but the Michiganders were
blissfully unaware of this so I had some
lines and I that were produced by women
from Detroit I played that in four
Michiganders and I said I’m creating a
speech speech synthesizer help me pick
out the vowels that you hear if they
thought she was from Michigan they did
horrible why did they do horribly
because Michiganders think that they are
producing speech the way this was done
back in the day Walter Cronkite talks or
the newscasters talk when in fact
there’s a huge violent that’s happening
in the northern cities Cleveland Detroit
Chicago places like that they don’t
realize that this has taken place if
from the other hand I lie to my
respondents and say this woman is from
Windsor Ontario then they hear exactly
what it is that she’s doing so their
belief that Michiganders use perfectly
correct English overrides the acoustics
I call this filtering out some of the
acoustic signal the same thing
unfortunately happens with our speech
therapists one of the things that can
happen if you have had a cleft palate as
you can have what’s called hyper
nasality and it turns out that if you
show a picture of a client to a speech
therapist and say diagnose the degree of
hyper nasality in this person they’ll
rate it as higher if they think the
person has a cleft palate versus if they
don’t I actually have a student right
now that’s working on now augmenting
what speech therapists do
actually speech recognition because
that’s one place that speech recognition
can help it can help make sure that it’s
an objective measure of hyper nasality
versus subjective all right now what
happens with the Michiganders they’re
just out of luck they’re just gonna
bliss way to go I’m thinking they also
I’m like Walter Cronkite or Justin
Bieber or something like that so just to
finish up here we have this big great
cerebral cortex that we just use it why
do we use it well we listen to speech in
noisy environments all of the acoustics
that are happening from the lights from
the air-conditioning can mask some of
the things coming out of my mouth so
what happens is we use other things we
use context we use vision we use
everything in order to help us make
sense of what our interlocutors are
saying we listen to people with all
different dialects so we have to kind of
have some ideas about how these dialects
work we listen to people who don’t
perform speech necessarily the way that
the dictionary says the Deuter that they
were taught from their English teachers
so we have to have a way of making sense
of that and what we do is we engage our
entire cerebral cortex so that if I say
the end or the end or the end every
child knows that the last one means that
the talk is over but the sidewalk with
the automated speech recognition system
in their head shut down 20 seconds ago
thank you very much
[Applause]
[Music]
[Music]
or it is basically just a hole by their
boot events so what’s happening right
now is I’m manipulating the shape of my
multi-track and that’s manipulating
physics and having a whole bunch of
acoustic events and you’re interpreting
that at speech so one of the things that
you could say happens in reception which
is just you’re suing acoustic pattern
matching you’re taking the frequencies
that are coming out of my mouth your
magic
you’re saying I understand what it is
that she’s saying and there’s a lot of
evidence that that is what humans do we
have auditory canal we have a bow chain
because amplifies frequencies coming out
of my mouth we have a cochlea that has
hair cells that librarians sympathy to
those frequencies that it’s connected to
some auditory nerves that tells our
brain which frequencies are being
created and magically somehow if you
hear a 300 Hertz and a 21-13 frequencies
that you say she was produce right so we
could say we’re just matching acoustics
and there’s no reason like Siri can’t do
that there’s no reason why Alexa can do
that speech recognition should be a
breeze well it turns out there’s a lot
more that goes into it and that’s what
I’ll try to convince you of today so
we’ll start with this quick little
experiment don’t let the sky play again
and what is it what is it that you’re
hearing
ah alright stop for one sec alright
we’re gonna play it one more time but
what I want you to do this time just
close your eyes all right close your
eyes and here we go again okay you can
open your eyes closed doing acoustic
pattern matching you’re taking the
acoustics and you’re brainless saying
match those to a pattern
I hear ba but when your eyes were open
what was happening was you are also
taking visual cues in in a nutshell this
is called the McGurk effect and in a
nutshell this is just your visual cortex
messing with your audio cortex so what
I’ll do for the rest of the day is try
over the rest of my luck not the rest of
the day you only have to say is to try
to convince you that when the whole
cerebral cortex gets involved it’s a
mess it’s a big mess but it’s a useful
mess all right so let’s go on to this
next thing this is it easy what’s the
past tense of how do you make a verb
past tense in English
Edie of course Edie what about spoken
English what do you add yeah duck right
hug
past tense hug what do you add to kiss
are you adding a D hearing a T right
what are your adding to Wade you’re any
good so you have two all right so it’s
not Edie it’s actually at least three
it’s actually a couple more but we’re
gonna stick with these three right now
so maybe what’s happening is you
memorize
in the English language and you memorize
every single past tense that goes with
those particular verbs but if I give you
word you’ve never heard before each one
of you is going to say the exact same
thing about what the past tense of
English is so the past tense of myth is
mid mid sorry mid a sense of meet with
the tea past tense of duck does it
alright you know do this alright so this
is something that you know about
language that you didn’t even know that
you knew and let’s do you think this is
a little weird thing with the past tense
same exact thing happens with the plural
you all know the plural of English in
English is to make a noun plural you
just add a s what about what do you add
to a cat s cats sure what about dogs
that’s a Z and in fact most of the time
we can make a noun plural you add a Z
right so you think it’s an ass but it’s
actually a Z and then to fish I don’t
want to hear fish is the plural of
fishes fish
and then you add right alright so yeah I
mean those right and then if I give you
words that you’ve never heard before
you’re all going to do the exact same
thing all right so the personal thing
that’s happening here is in our
conscious mind we’re sort of
oversimplifying things we’re saying at
several that we know that when we have
to add an alveolar to verb that ends a
voiced sound a daddy voices at a t if
it’s another IV later
know just make it good all right say
there’s a really set of complex rules
that tells us the plural no let’s just
simplify it and say that it’s s so one
of the things that our brain does is
takes this really complex system and
simplifies it what do we do with the
speech recognition system well we find a
linguist and we say help us with these
rules that are in people’s heads that
they don’t know that are there because
we’re going to need them to make our
automated speech recognition system work
well all right let’s do another little
test here I’m going to just say a phrase
what’s the phrase that I’m saying the
last speech so the last speech well
we’re at a TED talk so what you probably
heard me say was the last speech all
right but if we were driving down the
coast and I said this is the last speech
you wouldn’t hear it as speech you would
hear in his speech acoustically the last
speech and the last speech are exactly
the same they’re articulate orally the
same so their acoustic Li the same and
what you do is you use the context to
help you figure out whether you are
Beach
are you heard speech all right we do
this all the time can we build this into
a speech recognition system sure there
are probability trees or decision trees
that people are trying to build in the
system saying if these set of words
recognize then it’s more likely that
this word is this and not this so it’s
harder it’s more complex but we can
build that into the system all right
so you guys don’t have a problem with
this because you understand that context
is going to mean that I’m likely to say
the next speech if we’re here at a TED
talk and unless I’ve been nagging you
the whole way down our lovely Drive of
the coast you’re going to hear it as the
last beach and at the last speech all
right what else goes into this all right
in the 1950s people started using kind
of old-timey speech synthesis to do
experiments on the human perceptual
system one of the things that they did
was they realized that they could change
the precept of what vowel somebody was
hearing by manipulating a certain set of
things so one of the things that Varro’s
what vowels are basically just kind of
categorical boundaries in this big blob
of acoustic space so yeah that’s just
all a blob of a bunch of different
acoustics based on almost an infinite
amount of different acoustic patterns
are showing up but your brain says
divided the acoustic pie a certain way
but the frequencies that come out of my
vocal tract are going to be very
different from the frequencies that come
out of somebody else’s vocal tracts it’s
completely dependent on the size of your
mouth the size of your larynx how long
your vocal tract is what kind of what
your teeth are like all these kinds of
things affect the frequencies so somehow
we have to figure out a way as a human
perceiver to make sense of this and one
of the things it turns out that people
do is they pay attention to what the
range of frequencies are that are coming
out of their interlocutors mouth so they
synthesized with their old timey
synthesizer the word bet all right if
you have a vowel that sort of midway
between bit and bet you want to see what
people are going to perceive and it
turns out if you put this word in the
end of a sentence that has the top
frequencies people are going to say I
hear bet but if instead you have the
bottom set of frequencies people are
gonna say I hear bit in other words what
people do is they calibrate their
interlocutors vocal space they say
here’s the frequencies that I hear and
I’m going to contextualize all the
vowels that they come out based on what
I have calibrated for them there’s been
lots of studies since the 50s with much
better since speech synthesis showing
exactly this so one of the things you do
is you just say I hear somebody and now
I’m going to do some calculations and
use that information for how I interpret
everything else that they say we can
build this into a synthesizer – it’s a
little trickier but we can just have the
cyntha I means I’m sorry they a
recognizer we can have the recognizer
get the range of the pitch but somebody
is producing and we can work that into
our calculations do we do this at
another level sure one of the things
that happens is for most people males
have longer vocal tracks and longer
longer vocal cords than females what
they have found in a lot of different
studies is that just seeing a face that
looks female that people have identified
as female or people have identified as
male changes the precept of certain
things to give you one example the
difference between sip and ship so the
difference between and Shh
it’s just where that band of white noise
is first it’s in higher frequencies
first it’s in lower frequencies if I
synthesize a token that’s halfway
between sip and ship I can try to figure
out what it is people are going to do
with that and one of the things that
happens is if they think they’re
listening to a female they say oh that’s
ship
if they think the very same signal is of
coming out of a male mouth they say oh
that’s sip in other words even before
people have heard the frequencies coming
out of
their interlocutors mouths they’ve made
some decisions about what kind of things
will come out of that person’s mouth in
order in other words they sort of pre
calibrate and they do this with kids
because kids are gonna have even smaller
vocal tracts smaller vocal cords higher
frequencies they tend to look at
people’s height they tend to look at all
kinds of different things about their
interlocutor before they even open their
mouth all right can we build this into a
synthesizer sure I’m sorry I keep saying
synthesizer recognizer speech
recognition system can we build this in
yes but the ramifications of this have
been that for the first several decades
of work on automated speech recognition
male voices were recognized much better
than female voices leading one of my
colleagues to say in the early 90s I
don’t know what it is about women they
just can’t seem to be recognized by our
machines so in any case we fixed some of
this now so in the decades since it
started we fixed some of this alright
but here’s where it gets even more
tricky if you’re building an automated
speech recognition system we interact
with people with all kinds of different
dialects so to give you an example in
Houston some people have marched it and
ad in front of nasals in other words pin
and pin are homophones you stick with a
pin you write with a pen all right other
people have that as two separate phone
to separate lexical items pin pen all
right
who says pin pin who says pin pen well
one of the things we know is that for
Anglos older people are more likely to
say pin pin younger people are more
likely to say pin pen do people know
this well if you do a perceptual test
and give you something like rinse and
rinse will also be homonyms if you have
this merger they won’t if you don’t so
you’re either rinse and rinse or rinse
and rinse if I do a perceptual test
ask somebody to choose what they’re
hearing but I hooked them up to an eye
tracker one of the things I might be
interested in is where their eyes are
going well if I put the picture of the
older lady in the middle and say this is
who you’re listening to it turns out
that people’s eyes sort of move back and
forth between rinse and rinse if on the
other hand I put the younger picture in
there their eyes go right to prints in
other words you might not have known
this at a conscious level but people
react as if this information is in their
heads I have an idea that older people
are likely older white people are more
likely to emerge this and younger white
people are more likely to not so I’m
going to use that information all right
what do we do with this with this with
an recognizer
I’ll talk about that in just a second
but let me give you one more example
what else can we use well there is a
feature of almost every dialect of
American English that says instead of
saying alveolar stop at the end of the
word I can replace that with a glottal
stop so instead of saying what I can say
what so I didn’t break my time up to my
alveolar Ridge in the second one and the
first one I went what and bring brought
my tongue to the my alveolar Ridge in
the second one what I just closed my
glottis but you probably all interpreted
the second one as what so I guess all of
the people of American English got
together and said what how about if we
just say that the grotto can stand in
for the alveolar if it’s a teeth okay
[Music]
and there are some African American
that stayed back and said you know what
D is also an alveolar so how about if we
make not only the T stand in for the
alveolar offer they got up with the
glottal stop but also D so good is
interpreted as good but so is good if I
stop things at the alveolar Ridge I mean
at my glottis that can still stand in
for a D let’s do that though just taking
this same rule just pushing it one step
farther okay that sounds good should we
tell the white people for the Anglos
that can only be SCT for the
african-americans that could be SCT or
sa ID if we put the dude on the right in
the middle and say this is what you’re
listening to people pick set if we have
the dude on the left and say this is who
you’re listening to people pick either
said or set alright so this is probably
a rule that you’re not aware you if I
ask you what people do with alveolar
stops if I ask you what the difference
is between African American and European
European American English you might not
say this but it turns out people know
this at some level so what do we do with
our speech recognition system we kind of
say this is there’s no way we can really
build this into the model except to say
some people might do this and some
people might do this all right
so perceived age of the speaker
influences how we perceive things
pristine ethnicity so does gender lines
on a map and perceive disability and
those
three last ones are in red because we’ve
done some experiments showing that in
fact those things can even override the
acoustic signal I did a lot of work with
my people I grew up with in Michigan
because one of the things that I found
interesting was a lot of Canadian things
that people were associating with Canada
were creeping into anger speech in
Michigan but the Michiganders were
blissfully unaware of this so I had some
lines and I that were produced by women
from Detroit I played that in four
Michiganders and I said I’m creating a
speech speech synthesizer help me pick
out the vowels that you hear if they
thought she was from Michigan they did
horrible why did they do horribly
because Michiganders think that they are
producing speech the way this was done
back in the day Walter Cronkite talks or
the newscasters talk when in fact
there’s a huge violent that’s happening
in the northern cities Cleveland Detroit
Chicago places like that they don’t
realize that this has taken place if
from the other hand I lie to my
respondents and say this woman is from
Windsor Ontario then they hear exactly
what it is that she’s doing so their
belief that Michiganders use perfectly
correct English overrides the acoustics
I call this filtering out some of the
acoustic signal the same thing
unfortunately happens with our speech
therapists one of the things that can
happen if you have had a cleft palate as
you can have what’s called hyper
nasality and it turns out that if you
show a picture of a client to a speech
therapist and say diagnose the degree of
hyper nasality in this person they’ll
rate it as higher if they think the
person has a cleft palate versus if they
don’t I actually have a student right
now that’s working on now augmenting
what speech therapists do
actually speech recognition because
that’s one place that speech recognition
can help it can help make sure that it’s
an objective measure of hyper nasality
versus subjective all right now what
happens with the Michiganders they’re
just out of luck they’re just gonna
bliss way to go I’m thinking they also
I’m like Walter Cronkite or Justin
Bieber or something like that so just to
finish up here we have this big great
cerebral cortex that we just use it why
do we use it well we listen to speech in
noisy environments all of the acoustics
that are happening from the lights from
the air-conditioning can mask some of
the things coming out of my mouth so
what happens is we use other things we
use context we use vision we use
everything in order to help us make
sense of what our interlocutors are
saying we listen to people with all
different dialects so we have to kind of
have some ideas about how these dialects
work we listen to people who don’t
perform speech necessarily the way that
the dictionary says the Deuter that they
were taught from their English teachers
so we have to have a way of making sense
of that and what we do is we engage our
entire cerebral cortex so that if I say
the end or the end or the end every
child knows that the last one means that
the talk is over but the sidewalk with
the automated speech recognition system
in their head shut down 20 seconds ago
thank you very much
[Applause]
[Music]
Please follow and like us: