1
00:00:00,167 --> 00:00:10,292
...

2
00:00:12,584 --> 00:00:14,167
- "ROSETTA" is the abbreviation

3
00:00:14,500 --> 00:00:17,292
of "Robot for Adapted
Subtitling and All Translations."

4
00:00:17,626 --> 00:00:20,250
This is a public-private research project

5
00:00:20,584 --> 00:00:23,459
labeled by Cap Digital
and co-financed by Bpifrance,

6
00:00:23,792 --> 00:00:26,250
as part of the Future Investments Program.

7
00:00:26,584 --> 00:00:29,751
The ambition of the ROSETTA project
is to make audiovisual content

8
00:00:30,042 --> 00:00:32,667
of all types more accessible to users.

9
00:00:32,999 --> 00:00:33,834
For that,

10
00:00:34,125 --> 00:00:37,209
ROSETTA develops systems
for the automatic production

11
00:00:37,542 --> 00:00:40,417
of multilingual and adapted subtitles.

12
00:00:40,751 --> 00:00:43,334
ROSETTA also offers
an exploratory translation

13
00:00:43,667 --> 00:00:46,542
into French sign language,

14
00:00:46,876 --> 00:00:50,250
using the animation
of a 3D avatar signing in LSF.

15
00:00:53,959 --> 00:00:55,250
This project is going on

16
00:00:55,584 --> 00:00:58,959
over 3 years and 6 months,
and it has 8 modules.

17
00:00:59,250 --> 00:01:02,375
Module 1 is transverse.
It coordinates project management and

18
00:01:02,709 --> 00:01:05,792
communication between
the different partners,

19
00:01:06,083 --> 00:01:08,709
private companies and public laboratories.

20
00:01:09,000 --> 00:01:11,292
Module 2 aims to collect
and produce the corpus

21
00:01:11,626 --> 00:01:13,959
that will serve for learning

22
00:01:14,250 --> 00:01:16,042
and evaluation of automatic models.

23
00:01:16,375 --> 00:01:19,918
Module 3 aims to produce,
through neural networks,

24
00:01:20,209 --> 00:01:23,667
French subtitles adapted
for broadcasting and corresponding

25
00:01:23,999 --> 00:01:25,542
to various audiences,

26
00:01:25,876 --> 00:01:27,375
including the deaf
and hard of hearing audience.

27
00:01:27,709 --> 00:01:30,375
The goal of Module 4 is
to automatically produce

28
00:01:30,709 --> 00:01:33,584
multilingual subtitles,
thanks to models

29
00:01:33,918 --> 00:01:37,167
of neural translation specialized
for the subtitling task

30
00:01:37,500 --> 00:01:40,250
from French to English, Spanish, Italian,

31
00:01:40,584 --> 00:01:41,667
Portuguese and Chinese.

32
00:01:42,667 --> 00:01:46,000
Module 5 is dedicated
to the generation of animation

33
00:01:46,334 --> 00:01:49,959
of a 3D avatar signing in LSF,
to experiment

34
00:01:50,250 --> 00:01:53,250
the possibilities of translating
directly into LSF

35
00:01:53,584 --> 00:01:55,667
the content of television newspapers.

36
00:01:55,999 --> 00:02:00,125
Module 6 aims to assess quality,
utility and usability

37
00:02:00,459 --> 00:02:02,959
of previous modules, from the perspective

38
00:02:03,250 --> 00:02:07,209
both of experts in translation in LSF
and production of subtitles

39
00:02:07,542 --> 00:02:10,292
and users with their specific needs.

40
00:02:10,626 --> 00:02:14,584
Module 7 attempts to anticipate
the success of the project,

41
00:02:14,918 --> 00:02:17,959
by questioning users
about their expectations

42
00:02:18,250 --> 00:02:19,500
and positioning the project

43
00:02:19,834 --> 00:02:21,999
in relation to
existing needs and practices.

44
00:02:22,292 --> 00:02:23,334
Module 8, finally,

45
00:02:23,667 --> 00:02:26,959
integrates all of the modules
into a demonstrator.

46
00:02:27,250 --> 00:02:28,250
It promotes the project

47
00:02:28,584 --> 00:02:30,500
according to
a common dissemination strategy

48
00:02:30,834 --> 00:02:32,751
and accompanies the project's achievements

49
00:02:33,042 --> 00:02:34,542
in their operating phase.

50
00:02:39,584 --> 00:02:41,334
ROSETTA relies on recent developments

51
00:02:41,667 --> 00:02:44,959
of artificial intelligence,
particularly "deep learning"

52
00:02:45,250 --> 00:02:46,959
applied to automatic translation.

53
00:02:47,250 --> 00:02:50,292
And this, in different languages,
and for areas as as varied

54
00:02:50,626 --> 00:02:53,042
as entertainment, news
or learning.

55
00:02:54,417 --> 00:02:57,292
ROSETTA also relies
on the movement recording

56
00:02:57,626 --> 00:03:01,125
by Motion Capture and
a multilinear representation

57
00:03:01,459 --> 00:03:03,876
of sign language. These tools allow

58
00:03:04,167 --> 00:03:06,375
to do experiments beyond subtitles,

59
00:03:06,709 --> 00:03:08,334
in order to propose one day

60
00:03:08,667 --> 00:03:12,375
access to news content
directly in LSF.

61
00:03:17,751 --> 00:03:20,292
ROSETTA addresses 3 direct targets.

62
00:03:20,626 --> 00:03:22,292
1) Final users:

63
00:03:22,626 --> 00:03:27,792
deaf people, hard of hearing people,
people learning a foreign language

64
00:03:28,083 --> 00:03:31,626
or wanting to access content
in a language they master.

65
00:03:31,959 --> 00:03:35,250
2) Corporations and organisms
of production and distribution

66
00:03:35,584 --> 00:03:37,918
of digital content, particularly

67
00:03:38,209 --> 00:03:40,709
those who have an incentive
to produce adapted subtitles

68
00:03:41,000 --> 00:03:42,626
for much of their content.

69
00:03:42,959 --> 00:03:45,876
3) Administrations,
communication services,

70
00:03:46,167 --> 00:03:48,751
media, training organisms

71
00:03:49,042 --> 00:03:53,417
and any audiovisual content producer
wanting to make it more accessible.

72
00:03:58,709 --> 00:04:00,751
ROSETTA's prospects
are of 3 kinds.

73
00:04:01,125 --> 00:04:05,000
Provide greater access to
audiovisual content of all types

74
00:04:05,334 --> 00:04:06,834
to a wider population.

75
00:04:07,125 --> 00:04:08,000
Then,

76
00:04:08,334 --> 00:04:13,834
Automate the production chain and increase
the productivity of audiovisual players.

77
00:04:14,125 --> 00:04:15,000
Finally,

78
00:04:15,334 --> 00:04:20,125
Understand the transformation of
the accessibility market internationally.

79
00:04:21,542 --> 00:04:26,542
SYSTRAN