1
00:00:00,000 --> 00:00:05,640
Will artificial intelligence make language learning obsolete? If in ten

2
00:00:05,640 --> 00:00:11,190
years we will have a device that will allow us to communicate with people who speak other languages

3
00:00:11,190 --> 00:00:19,140
efficiently and without obstacles, will there still be people who will learn languages? This is

4
00:00:19,140 --> 00:00:24,480
a question that I often ask myself, which leads me to have existential crises because I think: "Oh my god,

5
00:00:24,480 --> 00:00:30,360
in ten years my work will no longer make sense! Nobody will learn Italian!" Now,

6
00:00:30,360 --> 00:00:36,240
no one can know how things are going to go, but I discussed it with my dad in my last episode

7
00:00:36,240 --> 00:00:41,220
of the podcast and why with my dad, why my dad (who just retired)

8
00:00:41,220 --> 00:00:47,610
has been involved with his entire career in intelligence. artificial and specifically

9
00:00:47,610 --> 00:00:53,112
voice recognition. You know when you talk to your devices, if you do?

10
00:00:53,112 --> 00:00:57,030
(automatic voice) In English it is "subscribe to Italian podcasts". Now

11
00:00:57,030 --> 00:00:59,280
I can act as an interpreter in a foreign language.

12
00:00:59,280 --> 00:01:00,330
Q: Hi, how are you doing?

13
00:01:00,330 --> 00:01:02,877
A: Hi, how are you? D: I'm fine, and you?

14
00:01:02,877 --> 00:01:05,790
A: I'm fine, and you? Q: I'm fine because I

15
00:01:05,790 --> 00:01:08,021
haven't signed up for Podcast Italiano yet. A: I'm sick because I haven't

16
00:01:08,021 --> 00:01:14,340
subscribed to an italian podcast yet. Q: What's an italian podcast? Do you mean 'Podcast

17
00:01:14,340 --> 00:01:15,840
Italiano', by any chance?

18
00:01:15,840 --> 00:01:20,040
A: What is an Italian podcast? Do you mean by any chance 'Italian Podcast'?

19
00:01:20,040 --> 00:01:23,340
D: Yes, of course, Italian Podcast, on YouTube! A: Yes, of course italian podcast on youtube.

20
00:01:23,340 --> 00:01:27,390
D: Of course, but do you know there's a podcast version of it as well?

21
00:01:27,390 --> 00:01:34,410
A: Sure, but did you know that there is also a podcast version?

22
00:01:34,410 --> 00:01:40,320
This technology already exists, maybe it's not perfect but in ten years? I do n't

23
00:01:40,320 --> 00:01:45,300
know ... Anyway my father was involved in speech recognition, that is the part

24
00:01:45,300 --> 00:01:52,560
of understanding by the human language machine. We did two episodes,

25
00:01:52,560 --> 00:01:56,850
the first about his career in the world of artificial intelligence and the

26
00:01:56,850 --> 00:02:02,250
second specifically about the linguistic applications of artificial intelligence

27
00:02:02,250 --> 00:02:09,570
and neural networks, which are so fashionable today. I leave you now an excerpt from the

28
00:02:09,570 --> 00:02:14,310
second episode in which we talk about these things. I hope you like it.

29
00:02:14,310 --> 00:02:22,080
Q: I wanted to talk now a little bit about the ... artificial intelligence in the linguistic

30
00:02:22,080 --> 00:02:28,170
field , so it seems to me that there are three uses mainly: translation,

31
00:02:28,170 --> 00:02:37,230
voice recognition and voice synthesis, so to make a machine speak, right ?

32
00:02:37,230 --> 00:02:38,910
A: Exactly. Q: Here,

33
00:02:38,910 --> 00:02:44,490
I wanted to start with the very speech recognition you've been dealing with for a

34
00:02:44,490 --> 00:02:53,160
long time and ask you how it works. You told me that once upon a time they just put linguistic

35
00:02:53,160 --> 00:02:59,400
knowledge in the machine and then this approach was completely abandoned.

36
00:02:59,400 --> 00:03:07,890
A: Exactly. Until, let's say, all the 90s, until 2000 even the automatic recognition

37
00:03:07,890 --> 00:03:16,770
systems had the skills of human experts inside. There was, for example, the phonetic knowledge

38
00:03:16,770 --> 00:03:23,340
of the language, that is, what are the basic sounds of the language and how they are organized among

39
00:03:23,340 --> 00:03:30,600
them; lexical knowledge, that is, how these sounds form words; then there was the

40
00:03:30,600 --> 00:03:39,060
syntactic knowledge, that is how words form correct sentences of the language, and these ...

41
00:03:39,060 --> 00:03:45,840
these knowledge were introduced by human experts, by phoneticians, linguists

42
00:03:45,840 --> 00:03:51,920
who inserted into the code (or in any case in the knowledge of the computer ) these informations.

43
00:03:51,920 --> 00:03:54,682
D: That is, in short, all the grammar of a language because it is grammar ...

44
00:03:54,682 --> 00:03:58,920
A: Yes, we spoke, in fact, of grammar, even the grammars of a language

45
00:03:58,920 --> 00:04:03,300
were inserted into ... into the computer. Q: So there were people with a

46
00:04:03,300 --> 00:04:09,460
grammar book that translated rules into computer instructions?

47
00:04:09,460 --> 00:04:14,380
A: Yes yes, exactly. At the beginning I myself had brought Italian grammar to the office.

48
00:04:14,380 --> 00:04:18,190
D: Of Serianni? A: The one I had in high school, I don't know whose it was,

49
00:04:18,190 --> 00:04:24,040
but and this went on for a very long time, then they ... they started using

50
00:04:24,040 --> 00:04:31,030
statistical methods too, at least for the lower level part of the sound. . But then lately,

51
00:04:31,030 --> 00:04:40,690
so I would say from 2013 onwards all this has literally disappeared, in the sense that

52
00:04:40,690 --> 00:04:47,860
models of neural networks called end to end models have arrived , i.e. models that go from start to

53
00:04:47,860 --> 00:04:55,570
finish, and these models in speech recognition start from signal that comes out of the microphone, from

54
00:04:55,570 --> 00:05:03,845
the waveform and reaches the words. Then a waveform enters, a sequence of words comes out.

55
00:05:03,845 --> 00:05:13,920
D: And so ... so everything in between, syntax, phonetics, morphology, vocabulary, everything ...

56
00:05:13,920 --> 00:05:19,080
everything in between happens magically? A: It happens magically in the interaction between these

57
00:05:19,080 --> 00:05:25,560
neurons. These end-to-end models are even more complicated than I have described

58
00:05:25,560 --> 00:05:34,620
neural networks . And yet all these ... this human knowledge and linguistic knowledge have disappeared.

59
00:05:34,620 --> 00:05:39,840
Oh my God, maybe they are present in the neural network but it is an opaque model, it is a so-called

60
00:05:39,840 --> 00:05:46,080
black box, it is a black box and therefore we do not know if the neural network has used them, has

61
00:05:46,080 --> 00:05:50,550
not used them. Will he have ... in his learning he will have rediscovered phonetics, will he have rediscovered

62
00:05:50,550 --> 00:05:56,580
linguistics? We don't really know. Q: So there is no way to understand what he is

63
00:05:56,580 --> 00:06:03,180
learning and what he "thinks", in quotation marks? A: No, I would say ... I would say no. This is

64
00:06:03,180 --> 00:06:09,150
perhaps one of the limitations of these neural networks that ... which perhaps is also a limitation of

65
00:06:09,150 --> 00:06:19,110
biological neural networks , in the sense that they are not inspectable. D: So they're not ... they're not very transparent.

66
00:06:19,110 --> 00:06:22,230
And so the only thing you need for these models

67
00:06:22,230 --> 00:06:27,870
is audio that is transcribed, right? A: Transcribed into words. And it takes

68
00:06:27,870 --> 00:06:36,090
many hours, we are talking about thousands of hours of transcribed recordings and the more there is, the

69
00:06:36,090 --> 00:06:45,120
better it will work. But it also takes days and days of very powerful computer computing

70
00:06:45,120 --> 00:06:51,510
to train the neural network, but eventually this network begins to figure out how to relate

71
00:06:51,510 --> 00:06:56,190
this strange input coming out of the microphone with words. In the case of recognition;

72
00:06:56,190 --> 00:07:00,570
in the case of translation it correlates the words in one language with the words in the other, even with

73
00:07:00,570 --> 00:07:06,150
totally different characters, this does not matter. D: Sure. And instead in the case of synthesis,

74
00:07:06,150 --> 00:07:11,520
that is, making the machine speak, how does that work? The correlation between what is it?

75
00:07:11,520 --> 00:07:17,730
A: Yes, it is exactly the other way around. The examples are pairs where the input is a sequence

76
00:07:17,730 --> 00:07:23,535
of words and the output is a waveform. Q: But is the ... waveform created from scratch?

77
00:07:23,535 --> 00:07:25,050
A: From scratch. D: Because once upon a time,

78
00:07:25,050 --> 00:07:30,510
perhaps you told me, they used just blocks of words, bits of

79
00:07:30,510 --> 00:07:34,830
words that were recombined in various ways. A: Exactly, what was in the synthesis,

80
00:07:34,830 --> 00:07:42,540
let's say, the classic one with knowledge introduced by man in which, in fact, man classified

81
00:07:42,540 --> 00:07:48,510
many pieces of recording that were then concatenated, the so-called concatenative synthesis.

82
00:07:48,510 --> 00:07:51,270
D: What can still be heard, for example, on some trains ...

83
00:07:51,270 --> 00:07:54,315
A: Otherwise, but I would say that ... D: Is it still used a lot?

84
00:07:54,315 --> 00:07:58,620
A: Yes yes yes, in the Italian railway systems the synthesis is still that of

85
00:07:58,620 --> 00:08:03,420
the 90 '. In my opinion or Amazon's systems they are already used though.

86
00:08:03,420 --> 00:08:10,640
Q: So if we have a system that ... the first step is to recognize the voice, then from the sound to the

87
00:08:10,640 --> 00:08:17,990
text, then the translation that translates into another language, into another text then, translated,

88
00:08:17,990 --> 00:08:25,100
and then we have the synthesis that he reads the translation aloud, in fact we have ... we have an

89
00:08:25,100 --> 00:08:30,950
interpreter, we have an interpreter. So you think that maybe translators and interpreters, translators

90
00:08:30,950 --> 00:08:36,425
already now but interpreters in the future, will be at risk, for example, in conferences?

91
00:08:36,425 --> 00:08:41,870
A: Unfortunately for human translators and interpreters I think yes, it will happen,

92
00:08:41,870 --> 00:08:48,770
or at least it will greatly reduce the scope ... the possibilities of ... work. In the sense that the

93
00:08:48,770 --> 00:08:56,540
sooner things will disappear more, let's say, more routine. I believe that the translation of technical manuals

94
00:08:56,540 --> 00:09:01,790
or product manuals is already done almost entirely automatically, even though there are still some

95
00:09:01,790 --> 00:09:09,950
errors inside. And then gradually also the work of interpreting maybe ... Maybe the two alternatives will exist

96
00:09:09,950 --> 00:09:14,960
for some time, the automatic one, cheaper and less precise, and

97
00:09:14,960 --> 00:09:21,920
the human one, more accurate and more expensive. D: And I think that even now it is starting to be

98
00:09:21,920 --> 00:09:27,560
a problem for some, for some translators, perhaps because clearly if

99
00:09:27,560 --> 00:09:33,710
the translator only has to correct the work done by a machine the pay,

100
00:09:33,710 --> 00:09:39,824
the pay will be ... it will be lower. A: Of course, and perhaps even less beautiful work.

101
00:09:39,824 --> 00:09:45,410
D: But even less beautiful yes. Going back to the topic of languages, do you think that learning languages

102
00:09:45,410 --> 00:09:52,040
will still be relevant in 10 or 15 years? This is a question I often ask myself. I don't know if there will be

103
00:09:52,920 --> 00:09:59,730
a device that will allow us to communicate with people who speak different languages ​​or a

104
00:09:59,730 --> 00:10:07,260
chip in the brain but also something less futuristic, let's say. Will it still be relevant?

105
00:10:07,260 --> 00:10:15,000
A: But I think that at least for a long time it will not become obsolete, in the sense that one learns

106
00:10:15,000 --> 00:10:24,660
a language for many reasons, but surely one thing is to be able to speak and communicate in a language

107
00:10:24,660 --> 00:10:32,040
without any device, another thing is to always have a device in hand or a device

108
00:10:32,040 --> 00:10:38,610
that acts as a mediator. I imagine that for reasons or work, or tourism, or even a little

109
00:10:38,610 --> 00:10:46,080
occasional these devices will certainly be used. Or maybe even in certain conferences which

110
00:10:46,080 --> 00:10:52,590
are occasional meetings of people of different ... so many nationalities could be used. But if

111
00:10:52,590 --> 00:11:00,270
one wants, in fact, to learn a language also to enter the culture of a nation, of a country

112
00:11:00,960 --> 00:11:10,110
this will remain totally irreplaceable, unless, as you mentioned, to implant

113
00:11:10,110 --> 00:11:16,530
artificial neural networks in the brain, then they implant the memory expansion of the language,

114
00:11:16,530 --> 00:11:20,820
you buy it and they bring it to you, but this really in my opinion is a little too far.

115
00:11:20,820 --> 00:11:26,340
And this was the excerpt, I hope you enjoyed it. If you want to hear the whole episode or the

116
00:11:26,340 --> 00:11:31,500
two episodes we did find the link below. Oh and if you didn't know,

117
00:11:31,500 --> 00:11:38,550
yes I have a podcast for those who learn the Italian language and that's why my name is Italian podcast,

118
00:11:38,550 --> 00:11:42,480
I know it's a bit weird. But I think these episodes in particular can be of interest to everyone

119
00:11:42,480 --> 00:11:49,080
even if you are Italian, because in short, my father is an expert in the sector. So let me

120
00:11:49,080 --> 00:11:54,240
know what you think, would you learn a language if there was a technology like it already

121
00:11:54,240 --> 00:12:01,140
exists but much more efficient than this? Maybe integrated into a device in our brain,

122
00:12:01,140 --> 00:12:06,370
or into a device that I know, a little more efficient than a mobile phone,

123
00:12:06,370 --> 00:12:11,890
which still acts as a bit of an obstacle, gets in the way between me and another person? Or do

124
00:12:11,890 --> 00:12:17,260
you think that you will continue to learn the language perhaps out of love for a culture or because you

125
00:12:17,260 --> 00:12:22,180
like learning languages? And what do you think most people will do? Let me

126
00:12:22,180 --> 00:12:27,010
know what you think as you understand it didn't have a real video for this week but

127
00:12:27,010 --> 00:12:33,430
let's be back next week with our usual schedule. Until next time! Bye Bye.