0:00:01.280,0:00:05.120
hi i'm michelle cook

0:00:03.199,0:00:07.759
i'm a professor in the geosciences

0:00:05.120,0:00:10.000
department at umass amherst

0:00:07.759,0:00:11.840
celia child and i made this video on the

0:00:10.000,0:00:14.400
effectiveness of captions for your live

0:00:11.840,0:00:16.320
lectures and presentations

0:00:14.400,0:00:18.480
captions like the ones you're seeing

0:00:16.320,0:00:20.000
below make verbal material more

0:00:18.480,0:00:22.560
accessible

0:00:20.000,0:00:23.439
a lot of people benefit from captions

0:00:22.560,0:00:25.039
people who are

0:00:23.439,0:00:26.800
not yet fluent in english people who

0:00:25.039,0:00:27.519
have auditory processing disorder or

0:00:26.800,0:00:29.439
deafness

0:00:27.519,0:00:32.320
and even folks who don't have access to

0:00:29.439,0:00:34.559
audio in shared workspaces

0:00:32.320,0:00:36.320
a study of television viewers in the uk

0:00:34.559,0:00:38.559
showed that only 20 percent

0:00:36.320,0:00:41.040
of regular caption users were deaf or

0:00:38.559,0:00:43.360
hard of hearing

0:00:41.040,0:00:44.399
now as we decrease the audio volume of

0:00:43.360,0:00:46.800
this recording

0:00:44.399,0:00:50.000
you may find yourself relying more and

0:00:46.800,0:00:51.440
more on the captions

0:00:50.000,0:00:53.360
in addition to increasing the

0:00:51.440,0:00:55.520
accessibility of your presentations

0:00:53.360,0:00:57.199
studies have shown that adding subtitles

0:00:55.520,0:01:00.239
or captions also improves

0:00:57.199,0:01:03.039
english language learning

0:01:00.239,0:01:04.720
for recorded talks like this one we can

0:01:03.039,0:01:05.680
add and edit the captions to the

0:01:04.720,0:01:07.439
recording

0:01:05.680,0:01:09.439
but what about our live lectures and

0:01:07.439,0:01:11.439
presentations

0:01:09.439,0:01:12.720
since captions help all of our students

0:01:11.439,0:01:16.720
why don't we regularly

0:01:12.720,0:01:18.479
provide captioning of our live lectures

0:01:16.720,0:01:19.920
our resistance could be due to two

0:01:18.479,0:01:21.680
factors either

0:01:19.920,0:01:23.200
we don't know how to implement captions

0:01:21.680,0:01:24.720
in our presentations

0:01:23.200,0:01:27.840
or maybe we don't believe that the

0:01:24.720,0:01:30.000
captions will be accurate enough

0:01:27.840,0:01:32.320
so in this short video celia and i are

0:01:30.000,0:01:34.799
going to both show you how to implement

0:01:32.320,0:01:36.320
auto captions in google slides and

0:01:34.799,0:01:38.079
microsoft powerpoint

0:01:36.320,0:01:40.640
and we're also going to show you how

0:01:38.079,0:01:44.640
well the artificial intelligence based

0:01:40.640,0:01:46.720
ai based captions perform

0:01:44.640,0:01:48.399
so with that i'm going to change to

0:01:46.720,0:01:52.079
google slides

0:01:48.399,0:01:54.880
you see i hit the present button here

0:01:52.079,0:01:56.399
and then the caption button in the bar

0:01:54.880,0:01:58.479
here

0:01:56.399,0:01:59.520
and with that we start to see the live

0:01:58.479,0:02:02.079
auto captions

0:01:59.520,0:02:03.840
at the bottom of the screen you'll

0:02:02.079,0:02:04.880
notice that i've kept my video on while

0:02:03.840,0:02:06.479
screen sharing

0:02:04.880,0:02:08.319
this is another good practice to

0:02:06.479,0:02:10.879
increase accessibility for books who

0:02:08.319,0:02:10.879
speech read

0:02:11.360,0:02:15.520
the image here shows the variety of ways

0:02:13.360,0:02:17.280
that you can implement captioning

0:02:15.520,0:02:19.599
for live presentations you can hire a

0:02:17.280,0:02:21.440
captionist trained as stenography

0:02:19.599,0:02:24.000
those captionists can provide accurate

0:02:21.440,0:02:26.720
captions with word error rate

0:02:24.000,0:02:28.959
lower than one percent and this is the

0:02:26.720,0:02:31.599
best option for some situations such as

0:02:28.959,0:02:33.680
disability accommodation

0:02:31.599,0:02:35.599
for most lectures where captionists are

0:02:33.680,0:02:37.760
not required for accommodation

0:02:35.599,0:02:39.760
you can easily add captions and increase

0:02:37.760,0:02:42.319
the accessibility of your presentations

0:02:39.760,0:02:43.440
using the live auto caption built into

0:02:42.319,0:02:45.280
google slides

0:02:43.440,0:02:49.040
or the version of microsoft powerpoint

0:02:45.280,0:02:51.440
that is part of office 365.

0:02:49.040,0:02:53.920
now you've probably seen some really bad

0:02:51.440,0:02:56.400
captions like those on youtube

0:02:53.920,0:02:58.239
youtube's word error rate is often

0:02:56.400,0:03:00.239
greater than 20

0:02:58.239,0:03:02.159
we call these terrible captions

0:03:00.239,0:03:04.080
corruptions

0:03:02.159,0:03:05.920
craptions are typically produced by

0:03:04.080,0:03:08.480
out-of-date auto-captioning software

0:03:05.920,0:03:10.319
that doesn't use artificial intelligence

0:03:08.480,0:03:11.519
the captions you're seeing right now use

0:03:10.319,0:03:14.959
ai and have more

0:03:11.519,0:03:17.200
are typically more accurate than youtube

0:03:14.959,0:03:19.120
with that i'm going to pass this to

0:03:17.200,0:03:23.360
celia who will show you just how

0:03:19.120,0:03:23.360
accurate the live auto captions can be

0:03:24.640,0:03:29.920
hello i am celia child i'm a geology

0:03:27.440,0:03:31.519
major at bryn mawr college

0:03:29.920,0:03:33.120
i conducted this study by making

0:03:31.519,0:03:35.280
recordings of my own voice

0:03:33.120,0:03:36.879
reading two passages that were rich in

0:03:35.280,0:03:38.560
geologic jargon

0:03:36.879,0:03:40.799
and then i played that recording into

0:03:38.560,0:03:44.400
both google slides and powerpoint to see

0:03:40.799,0:03:46.480
how well the ai captions performed

0:03:44.400,0:03:48.879
first i'm going to show you how to set

0:03:46.480,0:03:51.440
up auto caption in powerpoint

0:03:48.879,0:03:53.040
you need to have office 365 which comes

0:03:51.440,0:03:55.439
with the subtitle option

0:03:53.040,0:03:58.000
under slideshow see you just click this

0:03:55.439,0:04:00.560
button always use subtitles

0:03:58.000,0:04:01.360
and underneath that there's a button you

0:04:00.560,0:04:03.439
can press

0:04:01.360,0:04:05.120
to choose from a variety of english

0:04:03.439,0:04:07.680
accents or even

0:04:05.120,0:04:10.959
speak non-english languages and have the

0:04:07.680,0:04:14.000
captions report in english

0:04:10.959,0:04:16.160
before i tested the recordings we made a

0:04:14.000,0:04:18.959
list of words from the passages that we

0:04:16.160,0:04:20.880
expected the software to miscaption

0:04:18.959,0:04:22.000
because they are unfamiliar geologic

0:04:20.880,0:04:24.000
jargon

0:04:22.000,0:04:25.919
as you can see in the second column the

0:04:24.000,0:04:27.759
performance of the software was actually

0:04:25.919,0:04:30.560
much better than we had originally

0:04:27.759,0:04:30.560
anticipated

0:04:32.400,0:04:36.720
we also measured word error rate to

0:04:34.560,0:04:38.720
assess the live auto caption under a

0:04:36.720,0:04:41.120
variety of conditions

0:04:38.720,0:04:42.479
using typical input sound quality

0:04:41.120,0:04:44.560
playing the recording into the

0:04:42.479,0:04:47.120
computer's built-in microphone

0:04:44.560,0:04:49.360
the ai-based caption software had about

0:04:47.120,0:04:51.759
a nine percent error rate

0:04:49.360,0:04:56.320
better than youtube craptions but not

0:04:51.759,0:04:58.720
nearly as good as a captionist then

0:04:56.320,0:04:59.680
i added background noise and reduced the

0:04:58.720,0:05:02.479
sound volume

0:04:59.680,0:05:04.320
and you can see with worse sound quality

0:05:02.479,0:05:07.520
comes much worse accuracy

0:05:04.320,0:05:07.520
in the auto captions

0:05:08.320,0:05:11.360
then i tried a high quality external

0:05:10.560,0:05:13.600
microphone

0:05:11.360,0:05:15.919
and speaking intentionally rather than

0:05:13.600,0:05:18.320
using the recording on my phone

0:05:15.919,0:05:20.400
under these best practices the error

0:05:18.320,0:05:22.240
rate was as low as 5

0:05:20.400,0:05:25.840
which is much better than anyone would

0:05:22.240,0:05:25.840
have expected

0:05:26.560,0:05:30.560
we also enlisted the help of friends

0:05:28.479,0:05:31.840
with a variety of non-native english

0:05:30.560,0:05:34.240
accents

0:05:31.840,0:05:35.919
the error rate varied greatly but was a

0:05:34.240,0:05:36.960
lot higher than the american accent

0:05:35.919,0:05:38.639
tests

0:05:36.960,0:05:40.560
we recommend that if you see the

0:05:38.639,0:05:42.639
captions frequently missing

0:05:40.560,0:05:43.840
some words with your accent that you

0:05:42.639,0:05:47.680
include those words

0:05:43.840,0:05:47.680
on your slide to help the audience

0:05:48.800,0:05:52.800
michelle here again to sum up we

0:05:51.680,0:05:55.520
recommend that you use

0:05:52.800,0:05:57.680
ai-based auto-captioning within all live

0:05:55.520,0:06:01.199
lectures and presentations

0:05:57.680,0:06:03.360
captions benefit everyone in your class

0:06:01.199,0:06:05.120
for best practices we recommend that you

0:06:03.360,0:06:07.840
use an external microphone

0:06:05.120,0:06:10.400
and speak deliberately and clearly when

0:06:07.840,0:06:13.360
we did that the ai-based auto caption

0:06:10.400,0:06:16.080
missed only a few words it performed

0:06:13.360,0:06:17.919
surprisingly well

0:06:16.080,0:06:20.000
we recommend that you practice with the

0:06:17.919,0:06:22.080
presentation software beforehand

0:06:20.000,0:06:24.160
and note the word that you are missed

0:06:22.080,0:06:25.120
with your particular accent and speech

0:06:24.160,0:06:27.360
patterns

0:06:25.120,0:06:29.280
you can add commonly missed words to the

0:06:27.360,0:06:32.880
slides to make the presentation more

0:06:29.280,0:06:32.880
accessible to your audience

0:06:32.960,0:06:37.360
and finally live auto caption is not a

0:06:35.440,0:06:39.199
substitute for high quality captions

0:06:37.360,0:06:41.360
requested to accommodate someone's

0:06:39.199,0:06:43.280
deafness or disability

0:06:41.360,0:06:44.880
ai based auto caption should never be

0:06:43.280,0:06:45.919
used as a cheaper alternative to

0:06:44.880,0:06:47.600
captionist

0:06:45.919,0:06:50.560
and request for captionists should

0:06:47.600,0:06:50.560
always be honored

0:06:50.960,0:06:55.120
you can read more about this project in

0:06:52.639,0:07:00.560
this 2020 eos article

0:06:55.120,0:07:00.560
thank you for your attention