00:00:07
You know how they say
00:00:08
there are two certainties
in life, right?
00:00:10
Death and taxes.
00:00:12
Can't we get rid of
one of those?
00:00:15
See, 100 years ago,
life expectancy was only 45,
00:00:17
can you believe that?
00:00:18
Then by the 1950s,
it was up to 65,
00:00:20
and today, it's almost 80.
00:00:22
Tomorrow, who knows? Right?
00:00:25
Healthcare has made
huge progress.
00:00:27
We've eradicated epidemics
that used to kill millions,
00:00:30
but life is fragile.
00:00:32
People still get sick,
or pass away
00:00:34
for reasons that maybe
00:00:36
should be, someday curable.
00:00:41
What if we could
improve diagnosis?
00:00:44
Innovate to predict illness
instead of just react to it?
00:00:47
In this episode,
00:00:49
we'll see how machine learning
is combating
00:00:51
one of the leading
causes of blindness,
00:00:53
and enabling a son
with a neurological disease
00:00:56
to communicate with his family.
00:00:58
AI is changing
the way we think
00:01:00
about mind and body,
00:01:01
life and death,
00:01:03
and what we value most,
00:01:05
our human experience.
00:01:07
[fanfare music playing]
00:01:09
[announcer] ...and our other
co-captain, Number 8!
Tim Shaw!
00:01:14
[crowd cheering]
00:01:18
[John Shaw] We've lived
with his football dream,
00:01:20
All the way back
to sixth grade
00:01:22
when his coach said,
00:01:23
"This kid is gonna go
a long way."
00:01:25
From that point on,
00:01:27
Tim was doing pushups
in his bedroom at night,
00:01:29
Tim was the first one
at practice.
00:01:32
Tim took it seriously.
00:01:34
[crowd screaming and cheering]
00:01:42
[whistle blows]
00:01:44
[announcer]
Number 8, Tim Shaw!
00:01:47
[crowd cheering]
00:01:49
I don't know what
they're doing out there,
00:01:51
and I don't know
who they comin' to!
00:01:53
[Robert Downey Jr.]
For as long as he can remember,
00:01:55
Tim Shaw dreamed
of three letters...
00:01:58
N-F-L.
00:02:00
[whistle blows]
00:02:01
He was a natural
from the beginning.
00:02:03
As a kid,
he was fast and athletic.
00:02:05
He grew into 235 pounds
of pure muscle,
00:02:09
and at 23,
he was drafted to the pros.
00:02:13
His dream was real.
00:02:16
He was playing
professional football.
00:02:19
[reporter] Hello,
I'm with Tim Shaw.
00:02:21
You get to start
this season right.
What's it feel like?
00:02:24
It's that amazing
pre-game electricity,
00:02:26
the butterflies are there,
and I'm ready to hit somebody.
You might wanna look out.
00:02:30
Hey, Titans fans,
it's Tim Shaw here,
00:02:32
linebacker
and special teams animal.
00:02:35
He loves what he does.
00:02:37
He says,
"They pay me to hit people!"
00:02:39
[crowd cheering]
00:02:42
I'm here to bring you
some truth,
00:02:44
a little bit of truth,
00:02:46
and so we'll call it
T-Shaw's truth,
00:02:47
'cause it's not
all the way true,
but it's my truth.
00:03:03
[Tim Shaw speaking]
00:03:27
[Tim from 2015 interview]
In 2012,
00:03:29
my body started to do things
it hadn't done before.
00:03:31
My muscles were twitching,
I was stumbling,
00:03:34
or I was not making a play
I would have always made.
00:03:39
I just wasn't
the same athlete,
00:03:40
I wasn't the same
football player
that I'd always been.
00:03:45
[Tim speaking]
00:04:03
[Downey] The three letters
that had defined Tim's life
up to that point
00:04:07
were not the three letters
that the doctor told him
that day.
00:04:10
A-L-S.
00:04:12
[Tim speaking]
00:04:37
Okay...
00:04:39
[Downey] A-L-S, which
stands for "amyotrophic
lateral sclerosis,"
00:04:43
is also known as
Lou Gehrig's Disease.
00:04:45
It causes the death of neurons
controlling voluntary muscles.
00:04:49
[Sharon Shaw] He can't even
scratch his head...
00:04:52
Better yet?
00:04:53
...none of those
physical things
00:04:55
that were
so easy for him before.
00:04:58
He has to think about
every step he takes.
00:05:01
So Tim's food comes
in this little container.
00:05:03
We're gonna
mix it with water.
00:05:06
[Tim speaking]
00:05:17
[Downey]
As the disease progresses,
00:05:19
muscles weaken.
00:05:20
Simple everyday actions,
like walking, talking,
and eating,
00:05:24
take tremendous effort.
00:05:49
Tim used to call me
on the phone in the night,
00:05:52
and he had voice recognition,
00:05:54
and he would speak to the phone,
00:05:56
and say, "Call Dad."
00:05:57
His phone didn't recognize
the word "Dad."
00:06:02
So, he had said to me...
00:06:05
[voice breaking] "Dad,
I've changed your name.
00:06:07
I'm calling...
I now call you "Yo-yo."
00:06:11
So he would say into his phone,
"Call Yo-yo."
00:06:16
[Sharon] Tim has stopped
a lot of his communication.
00:06:20
He just doesn't talk
as much as he used to,
00:06:23
and I, I miss that.
00:06:25
I miss it.
00:06:26
-What do you think
about my red beard?
-No opinion.
00:06:29
[snorts] That means
he likes it,
00:06:31
just doesn't wanna
say on camera.
00:06:32
Now, my favorite was when
you had the handlebar moustache.
00:06:35
[Downey] Language,
the ability to communicate
with one another.
00:06:39
It's something that makes us
uniquely human,
00:06:42
making communication
an impactful application
for AI.
00:06:47
[Sharon] Yeah, that'll be fun.
00:06:57
[Julie Cattiau]
My name is Julie.
00:06:59
I'm a product manager
here at Google.
00:07:01
For the past year or so,
I've been working on
Project Euphonia.
00:07:04
Project Euphonia
has two different goals.
00:07:06
One is to improve
speech recognition
00:07:09
for people who have a variety
of medical conditions.
00:07:12
The second goal
is to give people
their voice back,
00:07:15
which means actually recreating
the way they used to sound
00:07:18
before they were diagnosed.
00:07:20
If you think about
communication,
00:07:22
it starts with
understanding someone,
00:07:24
and then being understood,
00:07:26
and for a lot of people,
00:07:28
their voice is like
their identity.
00:07:31
[Downey]
In the US alone,
roughly one in ten people
00:07:34
suffer acquired
speech impairments,
00:07:36
which can be caused by
anything from ALS,
00:07:39
to strokes, to Parkinson's,
to brain injuries.
00:07:43
Solving it is a big challenge,
00:07:45
which is why Julie partnered
with a big thinker to help.
00:08:07
[Downey] Dimitri
is a world-class research
scientist and inventor.
00:08:11
He's worked at IBM,
Princeton, and now Google,
00:08:14
and holds over 150 patents.
00:08:18
Accomplishments aside,
00:08:19
communication
is very personal to him.
00:08:21
Dimitri has a pretty
strong Russian accent,
00:08:24
and also he learned English
when he was already deaf,
00:08:28
so he never heard himself
speak English.
00:08:32
Oh, you do? Oh, okay.
00:08:34
[Downey] Technology can't yet
help him hear his own voice.
00:08:37
He uses AI-powered
Live Transcribe
00:08:40
to help him communicate.
00:08:42
[Cattiau] Okay,
that's awesome.
00:08:43
So we partnered up with Dimitri
to train a recognizer
00:08:47
that did a much better job
at recognizing his voice.
00:08:50
The model that you're using
right now for recognition,
00:08:53
what data did you train it on?
00:09:01
[Downey] So, how does
speech recognition work?
00:09:07
First, the sound of our voice
is converted into a waveform,
00:09:10
which is really just
a picture of the sound.
00:09:13
Waveforms are then matched
to transcriptions,
00:09:16
or "labels" for each word.
00:09:18
These maps exist for most words
in the English language.
00:09:21
This is where
machine learning takes over.
00:09:24
Using millions
of voice samples,
00:09:26
a deep learning model
is trained
00:09:28
to map input sounds
to output words.
00:09:31
Then the algorithm uses rules,
such as grammar and syntax,
00:09:35
to predict each word
in a sentence.
00:09:37
This is how AI
can tell the difference
00:09:39
between "there," "their,"
and "they're."
00:09:42
[Cattiau] The speech
recognition model
that Google uses
00:09:45
works very well for people
00:09:47
who have a voice
that sounds similar
00:09:50
to the examples that were used
to train this model.
00:09:53
In 90% of cases,
it will recognize
what you want to say.
00:09:56
[Downey]
Dimitri's not in that 90%.
00:09:58
For someone like him,
it doesn't work at all.
00:10:01
So he created a model
based on a sample of one.
00:10:23
[Downey] But making
a new unique model
00:10:25
with unique data
for every new and unique person
00:10:27
is slow and inefficient.
00:10:29
Tim calls his dad "Yo-yo."
00:10:32
Others with ALS may call
their dads something else.
00:10:35
Can we build one machine
00:10:37
that recognizes
many different people,
00:10:40
and how can we do it fast?
00:10:42
[Cattiau] So this data
doesn't really exist.
00:10:45
We have to actually collect it.
00:10:46
So we started this partnership
with ALS TDI in Boston.
00:10:50
They helped us collect
voice samples
00:10:52
from people who have ALS.
00:10:54
This is for you, T. Shaw.
00:10:55
[all] One, two, three!
00:11:00
[all cheering]
00:11:02
I hereby accept your ALS
ice bucket challenge.
00:11:07
[yelping softly]
00:11:09
[Downey] When the
ice bucket challenge
went viral,
00:11:11
millions joined the fight,
and raised over $220 million
for ALS research.
00:11:17
There really is a straight line
from the ice bucket challenge
00:11:20
to the Euphonia Project.
00:11:29
ALS Therapy
Development Institute
is an organization
00:11:31
that's dedicated to finding
treatments and cures for ALS.
00:11:35
We are life-focused.
00:11:37
How can we use
technologies we have
00:11:40
to help these people right away?
00:11:42
Yeah, they're
actually noisier.
00:11:45
That's a good point.
00:11:47
I met Tim a few years ago
00:11:49
shortly after
he had been diagnosed.
00:11:51
Very difficult to go public,
00:11:53
but it was made
very clear to me
00:11:55
that the time was right.
00:11:57
He was trying to understand
what to expect in his life,
00:12:00
but he was also trying
to figure out,
00:12:02
"All right,
what part can I play?"
00:12:04
All the ice bucket challenges
and the awareness
00:12:07
have really inspired me also.
00:12:09
If we can just step back,
00:12:10
and say, "Where can I
shine a light?"
00:12:13
or "Where can I give a hand?"
00:12:15
When the ice bucket
challenge happened,
00:12:17
we had this huge influx
of resources of cash,
00:12:21
and that gave us the ability
00:12:23
to reach out to people with ALS
who are in our programs
00:12:26
to share their data with us.
00:12:28
That's what got us
the big enough data sets
00:12:31
to really attract Google.
00:12:36
[Downey] Fernando
didn't initially set out
00:12:38
to make speech recognition
work better,
00:12:40
but in the process of better
understanding the disease,
00:12:44
he built a huge database
of ALS voices,
00:12:47
which may help Tim
and many others.
00:12:54
[John] It automatically
uploaded it.
00:12:56
[Tim] Oh.
00:13:59
How many
have you done, Tim?
00:14:07
2066?
00:14:08
[Fernando Vieira] Tim,
he wants to find every way
that he can help.
00:14:12
It's inspiring to see
his level of enthusiasm,
00:14:16
and his willingness to record
lots and lots of voice samples.
00:14:20
[Downey] To turn
all this data into real help,
00:14:23
Fernando partnered
with one of the people
00:14:25
who started
the Euphonia Project,
Michael Brenner...
00:14:28
-Hey, Fernando.
-Hey, how are you doing?
00:14:30
[Downey] ...a Google
research scientist
00:14:32
and Harvard-trained
mathematician
00:14:34
who's using machine learning
00:14:35
to solve scientific
Hail Marys, like this one.
00:14:38
Tim Shaw has recorded
almost 2,000 utterances,
00:14:42
and so we decided
to apply our technology
00:14:44
to see if we could build
a recognizer that
understood him.
00:14:48
[Tim speaking]
00:14:51
The goal, right, for Tim,
is to get it so that it works
00:14:54
outside of the things
that he recorded.
00:14:56
The problem is
that we have no idea
00:14:58
how big of a set that
this will work on.
00:15:00
[Brenner] Dimitri had recorded
upwards of 15,000 sentences,
00:15:04
which is just an incredible
amount of data.
00:15:07
We couldn't possibly
expect anyone else
00:15:09
to record so many sentences,
00:15:11
so we know that
we have to be able to do this
00:15:14
with much less recordings
from a person.
00:15:16
So it's not clear it will work.
00:15:18
[Tim speaking]
00:15:22
-That didn't
work at all.
-Not at all.
00:15:24
He said, "I go
the opposite way,"
00:15:26
and it says,
"I know that was."
00:15:28
[Brenner]
When it doesn't recognize,
00:15:29
we jiggle around
the parameters of
the speech recognizer,
00:15:32
then we give it
another sentence,
00:15:34
and the idea is that
you'll get it to understand.
00:15:37
[Tim's recording]
Can we go to the beach?
00:15:42
-Yes! Got it.
-Got it.
00:15:43
That's so cool.
Okay, let's try another.
00:15:45
[Downey] If Tim Shaw
gets his voice back,
00:15:48
he may no longer feel
that he is defined,
00:15:50
or constrained,
by three letters,
00:15:53
but that's a big "if."
00:15:55
While Michael
and team Euphonia work away,
00:15:58
let's take a moment and imagine
what else is possible
00:16:01
in the realm of the senses.
00:16:03
Speech.
00:16:05
Hearing.
00:16:07
Sight.
00:16:08
Can AI predict blindness?
00:16:10
[truck horn beeps]
00:16:12
Or even prevent it?
00:16:50
[Downey] Santhi does not
have an easy life.
00:16:53
It's made more difficult
because she has diabetes,
00:16:56
which is affecting her vision.
00:17:05
[Downey] If Santhi doesn't
get medical help soon,
she may go blind.
00:17:12
[Dr. Jessica Mega]
Complications of diabetes
00:17:15
include heart disease,
00:17:16
kidney disease,
00:17:17
but one of the really
important complications
00:17:20
is diabetic retinopathy.
00:17:21
The reason it's
so important is that
00:17:24
it's one of the lead causes
of blindness worldwide.
00:17:27
This is particularly true
in India.
00:17:31
[giving instructions]
00:17:53
In the early stages,
it's symptomless,
00:17:55
but that's when it's treatable,
00:17:57
so you want to screen them
early on,
00:17:59
before they actually
lose vision.
00:18:05
In the early stages,
00:18:06
if a doctor is
examining the eye,
00:18:08
or you take a picture
of the back of the eye,
00:18:10
you will see lots of those
bleeding spots in the retina.
00:18:21
Today, the doctors are
not enough to do the screening.
00:18:25
We are very limited
ophthalmologists,
00:18:27
so there should be other ways
00:18:28
where you can screen
the diabetic patients
00:18:31
for diabetic complications.
00:18:32
[Downey] In the US,
00:18:33
there are about 74 eye doctors
for every million people.
00:18:37
In India, there are only 11.
00:18:39
So just keeping up with
the sheer number of patients,
00:18:42
let alone giving them
the attention and care
they need,
00:18:45
is overwhelming,
if not impossible.
00:18:48
[Dr. R. Kim] We probably see
about 2,000 to 2,500 patients
00:18:52
every single day.
00:18:53
[Mega]
The interesting thing
with diabetic retinopathy
00:18:56
is there are ways
to screen and get
ahead of the problem.
00:18:59
The challenge is that
not enough patients
undergo screening.
00:19:02
[Downey] Like Tim Shaw's
ALS speech recognizer,
00:19:05
this problem is also
about data,
00:19:08
or lack of it.
00:19:09
To prevent more people
from experiencing vision loss,
00:19:13
Dr. Kim wanted to
get ahead of the problem.
00:19:19
So there's a hemorrhage.
00:19:21
All these are exudates.
00:19:24
[Downey] Dr. Kim
called up a team at Google.
00:19:27
Made up of doctors
and engineers,
00:19:29
they're exploring ways
to use machine learning
00:19:31
to solve some of the world's
leading healthcare problems.
00:19:35
So we started with
could we train an AI model
00:19:39
that can somehow help
read these images,
00:19:41
that can decrease the number
of doctors required
00:19:45
to do this task.
00:19:47
So this is
the normal view.
00:19:48
When you start
looking more deeply,
00:19:50
then this can be
a microaneurysm, right?
00:19:53
-This one here?
-[man] Could be.
00:19:55
[Downey] The team
uses the same kind
of machine learning
00:19:58
that allows us
to organize our photos
00:20:00
or tag friends
on social media,
00:20:02
image recognition.
00:20:05
First, models are trained
00:20:07
using tagged images of things
like cats or dogs.
00:20:10
After looking at
thousands of examples,
00:20:12
the algorithm learns
to identify new images
00:20:15
without any human help.
00:20:17
For the retinopathy project,
00:20:19
over 100,000 eye scans
00:20:21
were graded
by eye doctors
00:20:23
who rated each eye scan
on a scale from one to five,
00:20:26
from healthy to diseased.
00:20:28
These images were then used
00:20:30
to train a machine
learning algorithm.
00:20:32
Over time,
the AI learned to predict
00:20:35
which eyes showed signs
of disease.
00:20:37
[Dr. Lily Peng]
This is the assistant's view
00:20:40
where the model's predictions
00:20:41
are actually projected
on the original image,
00:20:45
and it's picking up
the pathologies very nicely.
00:20:49
[Downey] To get help
implementing the technology,
00:20:51
Lily's team reached out
to Verily,
00:20:53
the life sciences unit
at Alphabet.
00:20:56
[Mega] So, how was India?
00:20:58
[Peng] Oh, amazing!
00:20:59
[Mega] Verily
came out of Google X,
00:21:01
and we sit at the intersection
of technology, life science,
and healthcare.
00:21:05
What we try to do is
think about big problems
00:21:07
that are affecting
many patients,
00:21:09
and how can we bring
the best tools
00:21:11
and best technologies
00:21:12
to get ahead of the problems.
00:21:14
The technical pieces
are so important,
and so is the methodology.
00:21:17
How do you capture
the right image,
00:21:19
and how does the algorithm work,
00:21:21
and how do you deploy
these tools not only here,
00:21:24
but in rural conditions?
00:21:25
If we can speed up
this diagnosis process
00:21:28
and augment the clinical care,
00:21:30
then we can prevent blindness.
00:21:32
[Downey] There aren't many
other bigger problems that
affect more patients.
00:21:36
Diabetes affects
400 million worldwide,
00:21:39
70 million in India alone,
00:21:41
which is why Jessica
and Lily's teams
00:21:44
began testing AI-enabled
eye scanners there,
00:21:47
in its most rural areas,
00:21:49
like Dr. Kim's
Aravind Eye Clinics.
00:21:52
-Is the camera on?
-Now it's on.
00:21:55
Yeah.
00:21:56
So once the camera is up,
00:21:57
we need to check
network connectivity.
00:21:59
[Sunny Virmani]
The patient comes in.
00:22:01
They get pictures
of the back of the eye.
00:22:03
One for the left eye,
and right eye.
00:22:05
The images are uploaded
to this algorithm,
00:22:08
and once the algorithm
performs its analysis,
00:22:10
it sends the results back
to the system,
00:22:13
along with
a referral recommendation.
00:22:15
It's good.
It's up and running.
00:22:17
Because the algorithm
works in real time,
00:22:20
you can get
a real-time answer to a doctor,
00:22:23
and that real-time answer
comes back to the patient.
00:22:30
[Kim] Once you have
the algorithm,
00:22:32
it's like taking
your weight measurement.
00:22:34
Within a few seconds,
00:22:36
the system tells you whether
you have retinopathy or not.
00:22:40
[Downey] In the past,
00:22:41
Santhi's condition could've
taken months to diagnose,
00:22:44
if diagnosed at all.
00:23:20
[Downey] By the time
an eye doctor would've
been able to see her,
00:23:23
Santhi's diabetes might have
caused her to go blind.
00:23:27
Now, with the help
of new technology,
00:23:29
it's immediate,
00:23:31
and she can take
the hour-long bus ride
00:23:33
to Dr. Kim's clinic in Madurai
00:23:35
for same-day treatment.
00:23:49
[Downey]
Now thousands of patients
00:23:51
who may have waited weeks
or months to be seen
00:23:53
can get the help they need
before it's too late.
00:24:11
Thank you, sir.
00:24:27
[Downey] Retinopathy
00:24:29
is when high blood sugar
damages the retina.
00:24:32
Blood leaks,
and the laser treatment
00:24:34
basically "welds"
the blood vessels
00:24:36
to stop the leakage.
00:24:38
Routine eye exams
can spot the problem early.
00:24:41
In rural or remote areas,
like here,
00:24:44
AI can step in and be
that early detection system.
00:24:57
[Pedro Domingos] I think
one of the most important
applications of AI.
00:25:01
is in places
where doctors are scarce.
00:25:04
In a way, what AI does
is make intelligence cheap,
00:25:08
and now imagine what
you can do when you make
intelligence cheap.
00:25:11
People can go
to doctors they
couldn't before.
00:25:14
It may not be
the impact that catches
the most headlines,
00:25:16
but in many ways it'll be
the most important impact.
00:25:31
[family chattering happily]
00:25:40
[Mega] AI now is
this next generation of tools
00:25:42
that we can apply to clinically
meaningful problems,
00:25:45
so AI really starts
to democratize healthcare.
00:25:56
[Mega] The work
with diabetic retinopathy
00:25:59
is opening our eyes
to so much potential.
00:26:03
Even within these images,
00:26:05
we're starting to see
some interesting signals
00:26:07
that might tell us about
someone's risk factors
for heart disease.
00:26:11
And from there,
you start to think about
00:26:13
all of the images that
we collect in medicine.
00:26:16
Can you use AI or an algorithm
00:26:18
to help patients and doctors
00:26:20
get ahead of a given diagnosis?
00:26:23
Take cancer as an example
of how AI can help save lives.
00:26:26
We could take a sample
of somebody's blood
00:26:29
and look for the minuscule
amounts of cancer DNA
00:26:31
or tumor DNA in that blood.
This is a great application
for machine learning.
00:26:36
[Downey]
And why stop there?
00:26:38
Could AI accomplish
00:26:40
what human researchers
have not yet been able to?
00:26:42
Figuring out how cells
work well enough
00:26:45
that you can understand
why a tumor grows
and how to stop it
00:26:48
without hurting
the surrounding cells.
00:26:50
[Downey] And if cancer
could be cured,
00:26:52
maybe mental health disorders,
00:26:54
like depression, or anxiety.
00:26:56
There are facial
and vocal biomarkers
00:26:58
of these
mental health disorders.
00:27:00
People check their phones
15 times an hour.
00:27:03
So that's an opportunity
00:27:05
to almost do, like,
a well-being checkpoint.
00:27:08
You can flag that
to the individual,
00:27:10
to a loved one,
00:27:11
or in some cases
even to a doctor.
00:27:14
[Bran Ferren]
If you look at the overall
field of medicine,
00:27:17
how do you do a great job
of diagnosing illness?
00:27:21
Having artificial intelligence,
00:27:24
the world's greatest
diagnostician, helps.
00:27:31
[Downey] At Google,
00:27:33
Julie and the Euphonia team
have been working for months
00:27:35
trying to find a way
for former NFL star Tim Shaw
00:27:38
to get his voice back.
00:27:39
[Dimitri Kanevsky speaking]
00:27:47
Yes! So Zach's team,
the DeepMind team,
00:27:50
has built a model
that can imitate your voice.
00:27:53
For Tim, we were lucky,
00:27:55
because, you know, Tim has
a career of NFL player,
00:27:59
so he did multiple
radio interviews
and TV interviews,
00:28:02
so he sent us
this footage.
00:28:04
Hey, this is Tim Shaw,
special teams animal.
00:28:07
Christmas is coming,
so we need to find out
00:28:10
what the Titans players
are doing.
00:28:12
If you gotta hesitate,
that's probably a "no."
00:28:14
[Cattiau] Tim will be able
to type what he wants,
00:28:17
and the prototype will say it
in Tim's voice.
00:28:21
I've always loved attention.
00:28:23
Don't know if you
know that about me.
00:28:24
[laughs] She's gonna
shave it for you.
00:28:26
[Downey] Interpreting speech
is one thing,
00:28:29
but re-creating the way
a real person sounds
00:28:31
is an order of magnitude harder.
00:28:33
Playing Tecmo Bowl,
eating Christmas cookies,
and turkey.
00:28:36
[Downey] Voice imitation
is also known as
voice synthesis,
00:28:40
which is basically
speech recognition in reverse.
00:28:43
First, machine learning
converts text back
into waveforms.
00:28:46
These waveforms are then
used to create sound.
00:28:50
This is how Alexa
and Google Home
are able to talk to us.
00:28:53
Now the teams from DeepMind
and Google AI
00:28:56
are working to create a model
00:28:58
to imitate the unique sound
of Tim's voice.
00:29:01
Looks like it's computing.
00:29:03
But it worked this morning?
00:29:05
We have to set expectations
quite low.
00:29:08
[Cattiau] I don't know how
our model is going to perform.
00:29:11
I hope that Tim will understand
00:29:13
and actually see the technology
for what it is,
00:29:16
which is a work in progress
and a research project.
00:29:20
[Downey] After six months
of waiting, Tim Shaw is
about to find out.
00:29:26
The team working on
his speech recognition model
00:29:28
is coming to his house
for a practice run.
00:29:33
[doorbell rings]
00:29:34
[dog barks]
00:29:39
[Sharon] Good girl,
come say hello.
00:29:41
-Hi!
-Oh, hi!
00:29:43
Welcome.
00:29:44
-Hi!
-Come in.
00:29:45
Thanks for having us.
00:29:47
[Sharon] He's met
some of you before,
right?
00:29:49
How are you
doing, Tim?
00:29:51
-Hi, Tim.
-Good to see you.
00:29:53
-Hello.
-Hello.
00:29:55
Hi.
00:29:56
Hi, I'm Julie.
We saw each other
on the camera.
00:30:01
It's warmer here
than it is in Boston.
00:30:03
[Sharon] As it should be.
00:30:05
[all laughing]
00:30:09
Okay.
00:30:11
Lead the way, Tim.
00:30:13
[Cattiau] I'm excited
to share with Tim
and his parents
00:30:16
what we've been working on.
00:30:18
I'm a little bit nervous.
I don't know if the app
00:30:21
is going to behave
the way we hope it will behave,
00:30:24
but I'm also very excited,
to learn new things
00:30:27
and to hear Tim's feedback.
00:30:29
So I brought
two versions with me.
00:30:32
I was supposed to pick,
but I decided
to just bring both
00:30:34
just in case one is better
than the other,
00:30:37
and, just so you know,
00:30:38
this one here was trained
00:30:41
only using recordings
of your voice,
00:30:44
and this one here was trained
using recordings of your voice,
00:30:47
and also from other participants
from ALS TDI
00:30:51
who went through
the same exercise of...
[laughing]
00:30:54
So, okay.
00:30:57
I was hoping we could
give them a try.
00:30:59
Are we ready?
00:31:03
Who are you talking about?
00:31:12
[app chimes]
00:31:15
It got it.
00:31:16
[John] It got it.
00:31:19
[gasps]
00:31:24
[Tim] Is he coming?
00:31:27
[app chimes]
00:31:29
Yes.
00:31:33
Are you working today?
00:31:36
[app chimes]
00:31:44
[chuckling]
00:31:46
It's wonderful.
00:31:48
[Cattiau] Cool.
00:31:49
Thank you
for trying this.
00:31:51
-Wow!
-It's fabulous.
00:31:53
[John] What I love,
it made mistakes,
00:31:55
-and then it corrected itself.
-Yeah.
00:31:57
I was watching it like,
"That's not it,"
00:31:59
and then it went...
[mimics app]
Then it does it right.
00:32:02
These were phrases,
00:32:04
part of the 70%
that we actually used
00:32:06
to train the model,
00:32:08
but we also set aside
30% of the phrases,
00:32:11
so this might not do as well,
00:32:14
but I was hoping that
we could try some of these too.
00:32:18
[John] So what
we've already done
00:32:20
is him using phrases
that were used
to train the app.
00:32:23
That's right.
00:32:24
Now we're trying to see
if it can recognize phrases
00:32:27
-that weren't part of that.
-[Cattiau] Yes, that's right.
00:32:30
So let's give it a try?
00:32:33
Do you want me to?
00:32:39
Do you have the time to play?
00:32:44
[app chimes]
00:32:47
What happens afterwards?
00:32:52
[app chimes]
00:32:53
Huh. So, on the last one,
00:32:55
this one got it,
and this one didn't.
00:32:58
-We'll pause it. So...
-I love the first one,
where it says,
00:33:02
-"Can you help me
take a shower?"
-[laughing]
00:33:04
-[Cattiau] That's not at all
what he said.
-[John] I know,
00:33:07
you've gotta be really careful
what you ask for.
00:33:10
[all laughing]
00:33:12
[John] So if, when it's
interpreting his voice,
00:33:15
and it makes some errors,
00:33:17
is there a way we can
correct it?
00:33:19
Yeah. We want to add
the option
00:33:22
for you guys to fix
the recordings,
00:33:24
but as of today,
because this is
the very first time
00:33:27
we actually tried this,
00:33:28
we don't have it yet.
00:33:30
[Cattiau] This is still
a work in progress.
00:33:32
We have
a speech recognition model
00:33:34
that works for Tim Shaw,
00:33:36
which is, you know, one person,
00:33:38
and we're really hoping
that, you know,
00:33:41
this technology can work
for many people.
00:33:43
There's something else
I want you to try,
00:33:46
if that's okay?
00:33:47
We're working with
another team at Google
called DeepMind.
00:33:50
They're specialized in
voice imitation and synthesis.
00:33:56
[Downey] In 2019,
00:33:57
Tim wrote a letter
to his younger self.
00:34:02
They are words written by
a 34-year-old man with ALS
00:34:05
who has trouble communicating
00:34:08
sent back in time
00:34:09
to a 22-year-old
on the cusp of NFL greatness.
00:34:15
[Cattiau] So let me
give this a try.
00:34:18
I just like using this letter
because it's just so beautiful,
00:34:21
so let me see
if this is gonna work.
00:34:27
[Tim's younger voice]
So, I've decided to
write you this letter
00:34:30
'cause I have so much
to tell you.
00:34:32
I want to explain to you
00:34:33
why it's so difficult
for me to speak,
00:34:36
the diagnosis, all of it,
00:34:38
and what my life is like now,
00:34:39
'cause one day,
you will be in my shoes,
00:34:41
living with the same struggles.
00:34:57
It's his voice,
that I'd forgotten.
00:35:14
We do.
00:35:47
[app chimes]
00:36:03
[app chimes]
00:36:05
We're so happy
to be working with you.
00:36:07
It's really an honor.
00:36:12
[John] The thought
that one day,
00:36:14
that can be linked with this,
00:36:18
and when you speak
as you are now,
00:36:20
it will sound
like that, is...
00:36:29
It's okay. We'll wait.
00:36:33
[Cattiau] There is
a lot of unknown
00:36:34
and still a lot of research
to be conducted.
00:36:37
We're really trying to have
a proof of concept first,
00:36:41
and then expand to not only
people who have ALS,
00:36:44
but people
who had a stroke,
00:36:45
or a traumatic
brain injury,
multiple sclerosis,
00:36:48
any types of
neurologic conditions.
00:36:51
Maybe other languages,
too, you know?
00:36:52
I would really like this
to work for French, for example.
00:36:56
[Mega] Wouldn't it be
a wonderful opportunity
00:36:58
to bring technology
to problems that we're solving
00:37:01
in life science and healthcare,
00:37:03
and in fact,
it's a missed opportunity
00:37:05
if we don't try to bring
the best technologies
00:37:07
to help people.
00:37:09
This is really
just the beginning.
00:37:10
[Downey]
Just the beginning indeed.
00:37:13
Imagine the possibilities.
00:37:15
I think in the imaginable
future for AI and healthcare
00:37:19
is that there is
no healthcare anymore,
00:37:22
because nobody needs it.
00:37:24
You could have an AI
that is directly talking
to your immune system,
00:37:27
and is actually preemptively
creating the antibodies
00:37:30
for the epidemics
that are coming your way,
00:37:32
and will not be stopped.
00:37:34
This will not happen
tomorrow, but it's
the long-term goal
00:37:37
that we can point towards.
00:37:43
[Downey] Tim had never
heard his own words
00:37:46
read out loud before today.
00:37:48
Neither had his parents.
00:37:51
[Tim] Every single day
is a struggle for me.
00:37:54
I can barely move my arms.
00:37:58
[John] Have fun.
00:38:00
I can't walk on my own,
00:38:02
so I recently started
using a wheelchair.
00:38:07
I have trouble
chewing and swallowing.
00:38:10
I'd kill for a good pork chop.
00:38:12
Yes, my body is failing,
but my mind is not giving up.
00:38:18
Find what's most important
in your life,
00:38:20
and live for that.
00:38:24
Don't let three letters, NFL,
00:38:26
define you...
00:38:28
[crowd cheering]
00:38:31
...the same way I refuse
to let three letters define me.
00:38:43
[John] One of the things
Tim has taught us,
00:38:45
and I think it's a lesson
for everyone...
00:38:48
Medically speaking,
Tim's life has an end to it.
00:38:51
In fact, five years ago
we were told he only had
two to five years left.
00:38:55
We're already past that.
00:38:58
He has learned very quickly
00:39:01
that today
is the day that we have,
00:39:05
and we can ruin today
00:39:08
by thinking about yesterday
00:39:10
and how wonderful it used to be,
00:39:12
and, "Oh, woe is me,"
and "I wish it was like that."
00:39:16
We can also ruin today
00:39:17
by looking into the future,
00:39:19
and in Tim's case,
00:39:20
how horrible
this is going to be.
00:39:22
"This is going to happen,"
00:39:23
"I won't be able
to do this anymore."
00:39:26
So if we go either
of those directions,
00:39:28
it spoils us
from being present today.
00:39:31
That's a lesson for all of us.
00:39:32
Whether we have
an ALS diagnosis or not,
00:39:35
try to see the good
and the blessing of every day.
00:39:38
You're here with us today.
00:39:40
It's going to be a good day.
