1
00:00:00,000 --> 00:00:02,240
Right now, somewhere in a sprawling call center

2
00:00:02,240 --> 00:00:04,599
in the Philippines, there's a worker speaking

3
00:00:04,599 --> 00:00:06,400
into a headset, right? They're trying to help

4
00:00:06,400 --> 00:00:08,320
a customer troubleshoot a router or something.

5
00:00:08,439 --> 00:00:10,720
Right. But the voice coming out of the other

6
00:00:10,720 --> 00:00:12,859
end of that line, the voice the customer actually

7
00:00:12,859 --> 00:00:15,740
hears, it belongs to a white American from the

8
00:00:15,740 --> 00:00:19,179
Midwest who literally does not exist. It's completely

9
00:00:19,179 --> 00:00:22,199
fabricated. Exactly. The worker's natural accent,

10
00:00:22,359 --> 00:00:25,359
their cadence, the actual physical reality of

11
00:00:25,359 --> 00:00:27,739
their vocal tract, it's all being captured and

12
00:00:27,739 --> 00:00:30,219
then just stripped away. It's mathematically

13
00:00:30,219 --> 00:00:33,740
replaced in real time by this synthetic digital

14
00:00:33,740 --> 00:00:36,939
shield. Which is just a profound manipulation

15
00:00:36,939 --> 00:00:39,000
of human interaction when you really stop to

16
00:00:39,000 --> 00:00:41,429
think about it. And what's wild here is that

17
00:00:41,429 --> 00:00:43,530
the technology doing this isn't some distant

18
00:00:43,530 --> 00:00:45,429
science fiction concept. Now it's happening right

19
00:00:45,429 --> 00:00:48,609
now. Right. It is a highly funded, rapidly scaling

20
00:00:48,609 --> 00:00:51,149
reality. And it's fundamentally altering how

21
00:00:51,149 --> 00:00:54,570
we even define human speech. So today we are

22
00:00:54,570 --> 00:00:57,850
plunging into this terrifying but honestly totally

23
00:00:57,850 --> 00:01:00,549
fascinating world of machine voices. We're doing

24
00:01:00,549 --> 00:01:03,909
a massive deep dive into Edward B. Kang's meticulous

25
00:01:03,909 --> 00:01:06,519
and Frankly, paradigm -shifting book. It's called

26
00:01:06,519 --> 00:01:09,739
Machine Voices, a Sonic Study of Artificial Intelligence.

27
00:01:09,959 --> 00:01:12,400
A really incredible piece of research. Oh, absolutely.

28
00:01:12,980 --> 00:01:15,060
So our mission today is to pull apart this text

29
00:01:15,060 --> 00:01:18,159
concept by concept for you. We are going to explore

30
00:01:18,159 --> 00:01:21,519
the hidden, often invisible, intersection between

31
00:01:21,519 --> 00:01:24,299
our human voices and the AI industry that is

32
00:01:24,299 --> 00:01:26,879
currently, well, devouring them. Devouring is

33
00:01:26,879 --> 00:01:28,959
the right word. And Kang's core premise here

34
00:01:28,959 --> 00:01:31,620
is incredibly compelling, because he starts with

35
00:01:31,620 --> 00:01:33,620
a contradiction. Like, I want you to just...

36
00:01:33,439 --> 00:01:35,540
Think about how we traditionally view the human

37
00:01:35,540 --> 00:01:38,299
voice. The French film theorist Michel Chian

38
00:01:38,299 --> 00:01:41,260
famously called the voice a strange object because

39
00:01:41,260 --> 00:01:43,840
it's intimate, right? It's elusive. You can't

40
00:01:43,840 --> 00:01:45,780
just hold it in your hand. Exactly. You can't

41
00:01:45,780 --> 00:01:48,359
touch it. It's deeply tied to our psychology,

42
00:01:48,659 --> 00:01:51,219
our physical bodies, our personal histories.

43
00:01:51,560 --> 00:01:53,700
It's the thing that makes you you. I mean, if

44
00:01:53,700 --> 00:01:55,620
you call your mother from an unknown number and

45
00:01:55,620 --> 00:01:57,459
just say one single word, she instantly knows

46
00:01:57,459 --> 00:02:01,049
it's you. Yes. But Kang's book exposes how the

47
00:02:01,049 --> 00:02:03,769
AI industry looks at this strange object and

48
00:02:03,769 --> 00:02:06,370
sees a completely different puzzle. They definitely

49
00:02:06,370 --> 00:02:08,569
don't see the intimacy. Not at all. They don't

50
00:02:08,569 --> 00:02:10,849
want the voice to be elusive or mysterious. They

51
00:02:10,849 --> 00:02:13,469
are actively trying to transform it into something

52
00:02:13,469 --> 00:02:16,750
manageable, something measurable, and above all

53
00:02:16,750 --> 00:02:19,710
else, something sellable. Right. And to pull

54
00:02:19,710 --> 00:02:21,830
off that transformation, to turn this intimate

55
00:02:21,830 --> 00:02:25,860
human breath into a literal commodity, Kang argues

56
00:02:25,860 --> 00:02:29,500
the AI industry relies on three derivative rationales.

57
00:02:29,659 --> 00:02:31,840
Yes. These are basically the foundational myths

58
00:02:31,840 --> 00:02:34,379
of the entire voice AI enterprise. Let's lay

59
00:02:34,379 --> 00:02:36,599
these out because they kind of dictate everything

60
00:02:36,599 --> 00:02:38,500
we're going to talk about today. What's the first

61
00:02:38,500 --> 00:02:42,300
one? So the first rationale is the belief that

62
00:02:42,300 --> 00:02:44,879
voices are knowable because they are mechanical.

63
00:02:45,479 --> 00:02:48,159
Meaning, if we just build the right machine,

64
00:02:48,240 --> 00:02:50,240
we can completely understand the human voice.

65
00:02:50,400 --> 00:02:53,500
OK. And the second? The second is kind of the

66
00:02:53,500 --> 00:02:57,199
inverse of that. It's that voices are inherently

67
00:02:57,199 --> 00:02:59,800
mechanical and therefore they are knowable. Like

68
00:02:59,800 --> 00:03:02,840
we aren't ghosts in a shell. We are just fleshy

69
00:03:02,840 --> 00:03:05,699
instruments producing acoustic data. Which naturally

70
00:03:05,699 --> 00:03:08,919
leads to the inevitable third step. Right. The

71
00:03:08,919 --> 00:03:11,919
ultimate conclusion. Because voices are both

72
00:03:11,919 --> 00:03:15,120
knowable and mechanical, they are replicable.

73
00:03:15,520 --> 00:03:17,840
If it's just a mechanism, we can just build a

74
00:03:17,840 --> 00:03:20,689
digital version of that exact mechanism. OK,

75
00:03:20,689 --> 00:03:22,750
let me pause here and push back on that a bit,

76
00:03:22,909 --> 00:03:25,370
or at least bring this down to earth for the

77
00:03:25,370 --> 00:03:27,370
person listening right now. Sure. Because if

78
00:03:27,370 --> 00:03:29,610
you are listening to this on a smartphone, the

79
00:03:29,610 --> 00:03:32,210
microphone you just used an hour ago to ask for

80
00:03:32,210 --> 00:03:34,689
driving directions has already mapped the physical

81
00:03:34,689 --> 00:03:37,110
geometry of your throat. Oh, absolutely. When

82
00:03:37,110 --> 00:03:40,349
you say, hey Siri, or use voice to text, you

83
00:03:40,349 --> 00:03:43,009
might think you're just using a handy tool. But

84
00:03:43,009 --> 00:03:46,210
under the surface, your specific laughs, your

85
00:03:46,210 --> 00:03:49,159
pauses, that unique timbre you have, It's all

86
00:03:49,159 --> 00:03:51,879
being mined as raw data to feed those exact three

87
00:03:51,879 --> 00:03:55,159
rationales. That is the exact reality Kang is

88
00:03:55,159 --> 00:03:57,919
analyzing here. We are moving from a paradigm

89
00:03:57,919 --> 00:04:01,439
where the voice is this phenomenon of human experience,

90
00:04:01,580 --> 00:04:04,860
this messy biological reality, to a paradigm

91
00:04:04,860 --> 00:04:07,319
where the voice is literally just a system of

92
00:04:07,319 --> 00:04:10,099
mechanisms. But getting to this point where an

93
00:04:10,099 --> 00:04:13,099
AI can clone a voice or an algorithm claims it

94
00:04:13,099 --> 00:04:14,819
can tell if you're a good employee just by the

95
00:04:14,819 --> 00:04:17,240
sound of your pitch, that obviously didn't happen

96
00:04:17,240 --> 00:04:19,829
overnight. No, it took decades of trial and error.

97
00:04:20,250 --> 00:04:22,870
Right. Kang traces this back to the very first

98
00:04:22,870 --> 00:04:25,569
incredibly clumsy attempts to visually capture

99
00:04:25,569 --> 00:04:28,970
the human voice. And that brings up this historical

100
00:04:28,970 --> 00:04:31,350
anecdote from the text that I think perfectly

101
00:04:31,350 --> 00:04:34,430
sets the stage for the modern AI obsession. You

102
00:04:34,430 --> 00:04:36,750
brought this up when we were prepping. Jacques

103
00:04:36,750 --> 00:04:40,449
de Volcanson's mechanical duck. Yes, the digesting

104
00:04:40,449 --> 00:04:43,550
duck. It's such a brilliant historical parallel

105
00:04:43,550 --> 00:04:46,189
that Kang uses. Set the scene for us. So we're

106
00:04:46,189 --> 00:04:48,930
talking about the 18th century. Vaucanson built

107
00:04:48,930 --> 00:04:52,550
this elaborate copper automaton duck. It could

108
00:04:52,550 --> 00:04:54,850
flap its wings, drink water, and supposedly,

109
00:04:55,110 --> 00:04:58,209
this was the big draw, it could eat grain, digest

110
00:04:58,209 --> 00:05:01,009
it, and then excrete it. And people were just

111
00:05:01,009 --> 00:05:03,569
mesmerized by this? Oh, completely. Okay. They

112
00:05:03,569 --> 00:05:05,589
thought he had mechanically replicated biology.

113
00:05:05,750 --> 00:05:08,170
They thought he built a digestive tract. They

114
00:05:08,170 --> 00:05:09,870
thought it was a total fraud, right? Completely

115
00:05:09,870 --> 00:05:12,149
fake. The grain went into one hidden compartment,

116
00:05:12,410 --> 00:05:15,490
and this preloaded fake excrement was released

117
00:05:15,490 --> 00:05:17,899
from a... Totally different compartment. Wow.

118
00:05:18,060 --> 00:05:19,600
Yeah, the machine wasn't digesting anything at

119
00:05:19,600 --> 00:05:21,579
all It was just mimicking the surface effect

120
00:05:21,579 --> 00:05:24,079
of digestion. Okay, so how does Kang connect

121
00:05:24,079 --> 00:05:27,879
this copper duck to modern AI? He uses it to

122
00:05:27,879 --> 00:05:30,639
ask a really critical question when a computer

123
00:05:30,639 --> 00:05:33,699
mimics a human voice perfectly has it actually

124
00:05:33,699 --> 00:05:35,939
replicated human intelligence or human biology

125
00:05:35,939 --> 00:05:40,060
or Is it just another mechanical duck? You know,

126
00:05:40,060 --> 00:05:42,480
is it just producing a highly convincing surface

127
00:05:42,480 --> 00:05:44,759
effect without any of the underlying reality?

128
00:05:44,939 --> 00:05:47,199
That is a phenomenal setup. Because the quest

129
00:05:47,199 --> 00:05:49,879
to map the underlying reality of the voice takes

130
00:05:49,879 --> 00:05:52,180
us straight into the FBI's historical obsession

131
00:05:52,180 --> 00:05:55,220
with something called voice prints. Yes, chapter

132
00:05:55,220 --> 00:05:57,779
one. Right. This is where the modern tech industry

133
00:05:57,779 --> 00:05:59,680
really started to grapple with mapping sound

134
00:05:59,680 --> 00:06:03,060
to identity. So we are in the 1960s and 70s,

135
00:06:03,180 --> 00:06:05,100
the FBI is desperate for new forensic tools.

136
00:06:05,540 --> 00:06:09,079
And there's this massive push to use speech spectrograms

137
00:06:09,079 --> 00:06:11,620
to identify criminals. Let's break down what

138
00:06:11,620 --> 00:06:13,420
a spectrogram actually is, just so we're all

139
00:06:13,420 --> 00:06:15,160
on the same page, because the mechanism here

140
00:06:15,160 --> 00:06:19,100
is crucial. Yeah, go for it. So, sound in the

141
00:06:19,100 --> 00:06:21,540
air is just changes in pressure, right? It's

142
00:06:21,540 --> 00:06:24,079
a wave. If you look at a standard audio file

143
00:06:24,079 --> 00:06:26,120
on your computer, you see a waveform. It just

144
00:06:26,120 --> 00:06:28,639
shows how loud the sound is over time. Right,

145
00:06:28,660 --> 00:06:30,399
it gets fatter when it's loud, skinnier when

146
00:06:30,399 --> 00:06:32,620
it's quiet, like a heart monitor. Exactly, like

147
00:06:32,620 --> 00:06:35,120
a heart monitor. But a waveform doesn't tell

148
00:06:35,120 --> 00:06:38,420
you what the sound actually is. A spectrogram

149
00:06:38,420 --> 00:06:42,100
is a much more complex visual map. It takes an

150
00:06:42,100 --> 00:06:44,920
audio signal and plots three things simultaneously.

151
00:06:45,680 --> 00:06:48,519
Time is on the horizontal axis, frequency or

152
00:06:48,519 --> 00:06:51,379
pitch is on the vertical axis, and the amplitude,

153
00:06:51,819 --> 00:06:55,000
the loudness, is represented by how dark or brightly

154
00:06:55,000 --> 00:06:58,319
colored the specific marks are. So it's basically

155
00:06:58,319 --> 00:07:00,000
taking a slice of audio and running it through

156
00:07:00,000 --> 00:07:02,560
a prism, right? Breaking it down into all its

157
00:07:02,560 --> 00:07:04,860
individual frequencies so you can literally look

158
00:07:04,860 --> 00:07:07,290
at the physical shape of a word. That's a really

159
00:07:07,290 --> 00:07:10,430
great way to visualize it. And in the 60s, this

160
00:07:10,430 --> 00:07:12,990
self -proclaimed expert named Lawrence Kursta

161
00:07:12,990 --> 00:07:15,769
looked at these visual shapes and made a massive,

162
00:07:15,889 --> 00:07:19,050
totally audacious claim. What did he claim? He

163
00:07:19,050 --> 00:07:21,649
heavily promoted the idea that a visual spectrogram

164
00:07:21,649 --> 00:07:24,649
of a voice was just as unique and just as reliable

165
00:07:24,649 --> 00:07:27,129
as a human fingerprint. Hence the term voice

166
00:07:27,129 --> 00:07:29,990
print. I mean, it's great marketing. It was brilliant

167
00:07:29,990 --> 00:07:32,660
marketing. Through these experiments he designed

168
00:07:32,660 --> 00:07:36,339
himself, Kirsta claimed an astounding 99 % accuracy

169
00:07:36,339 --> 00:07:38,740
rate in identifying speakers just by looking

170
00:07:38,740 --> 00:07:42,279
at these visual charts. 99 %? Yeah. And he was

171
00:07:42,279 --> 00:07:44,800
so confident, he set up his own company, VoicePrint

172
00:07:44,800 --> 00:07:47,699
Laboratories, and began offering a two -week

173
00:07:47,699 --> 00:07:50,259
training program to certify law enforcement officers

174
00:07:50,259 --> 00:07:52,680
as experts. OK, I have to stop you there. Two

175
00:07:52,680 --> 00:07:55,839
weeks? He thought he could train a cop to look

176
00:07:55,839 --> 00:07:58,899
at a complex frequency chart for two weeks? and

177
00:07:58,899 --> 00:08:01,860
achieve 99 % accuracy in identifying a human

178
00:08:01,860 --> 00:08:04,399
being. Yes. With enough certainty to literally

179
00:08:04,399 --> 00:08:07,279
send them to prison. He did. And for a while,

180
00:08:07,560 --> 00:08:09,600
courts were actually accepting this evidence.

181
00:08:09,740 --> 00:08:12,399
Juries were buying it. That's terrifying. It

182
00:08:12,399 --> 00:08:14,800
really is. But this is where the conflict between

183
00:08:14,800 --> 00:08:17,680
biology and machinery really explodes. Defense

184
00:08:17,680 --> 00:08:20,120
attorneys start pushing back hard. The scientific

185
00:08:20,120 --> 00:08:22,319
community started raising their eyebrows. It

186
00:08:22,319 --> 00:08:25,180
became a complete courtroom nightmare. So what

187
00:08:25,180 --> 00:08:27,800
happened? Well, in 1976, the FBI finally said,

188
00:08:27,819 --> 00:08:29,779
look, we need to know if this is real science.

189
00:08:30,079 --> 00:08:32,019
So they commissioned the National Academy of

190
00:08:32,019 --> 00:08:35,379
Sciences to investigate it. Right, the NAS. And

191
00:08:35,379 --> 00:08:37,179
they put together this heavy hitting committee.

192
00:08:37,620 --> 00:08:40,080
We're talking physicists, linguists, audiologists,

193
00:08:40,299 --> 00:08:43,080
lawyers. They spend years looking at Krista's

194
00:08:43,080 --> 00:08:46,220
methods. And in 1979, they release a massive

195
00:08:46,220 --> 00:08:48,720
report that essentially torches Krista's entire

196
00:08:48,720 --> 00:08:52,000
premise right down to the ground. And why they

197
00:08:52,000 --> 00:08:54,440
torched it is arguably the most important part

198
00:08:54,440 --> 00:08:57,549
of this whole story. Exactly. The committee pointed

199
00:08:57,549 --> 00:09:00,090
out a fundamental flaw in the metaphor of the

200
00:09:00,090 --> 00:09:02,529
voice print itself. Walk us through that flaw.

201
00:09:03,029 --> 00:09:04,809
Well, think about a traditional fingerprint.

202
00:09:05,230 --> 00:09:08,070
It's a direct graphical representation of human

203
00:09:08,070 --> 00:09:10,929
anatomy. It is the actual physical ridges on

204
00:09:10,929 --> 00:09:13,850
your skin pressing onto a surface. It's a direct

205
00:09:13,850 --> 00:09:16,809
one -to -one impression. But a voiceogram isn't

206
00:09:16,809 --> 00:09:20,710
that at all. It's indirect. Yes. It is an indirect

207
00:09:20,710 --> 00:09:24,590
representation of an incredibly complex, multi

208
00:09:24,590 --> 00:09:27,190
-layered process. Because when you speak, the

209
00:09:27,190 --> 00:09:29,769
resulting sound isn't just a static imprint of

210
00:09:29,769 --> 00:09:32,549
your vocal cords. It's heavily affected by your

211
00:09:32,549 --> 00:09:34,429
physiological state. Like, do you have a cold?

212
00:09:34,970 --> 00:09:37,370
It's affected by your psychological state. Are

213
00:09:37,370 --> 00:09:40,269
you stressed out during an interrogation? It's

214
00:09:40,269 --> 00:09:42,470
affected by the acoustic environment you're standing

215
00:09:42,470 --> 00:09:45,370
in. Is the room echoey? Is there traffic outside,

216
00:09:45,450 --> 00:09:48,309
right? Exactly. And crucially, it is profoundly

217
00:09:48,309 --> 00:09:50,830
altered by the mechanical equipment used to record

218
00:09:50,830 --> 00:09:53,289
it. A cheap mic sounds different than an expensive

219
00:09:53,289 --> 00:09:56,250
one. You simply cannot draw a straight objective

220
00:09:56,250 --> 00:09:59,549
line from a visual squiggle on a page to a person's

221
00:09:59,549 --> 00:10:02,149
unique biological identity. So treating a voice

222
00:10:02,149 --> 00:10:04,809
like a fingerprint is essentially like trying

223
00:10:04,809 --> 00:10:07,889
to read someone's soul through a footprint. But

224
00:10:07,889 --> 00:10:10,629
the footprint changes shape depending on if the

225
00:10:10,629 --> 00:10:13,149
person is tired or angry or carrying a heavy

226
00:10:13,149 --> 00:10:15,990
backpack or walking through mud. Right. And Kirsten

227
00:10:15,990 --> 00:10:17,830
was just looking at the mud and saying, yep,

228
00:10:18,009 --> 00:10:20,250
that's definitely Steve. That is a phenomenal

229
00:10:20,250 --> 00:10:22,529
analogy. The footprint is constantly shifting.

230
00:10:23,230 --> 00:10:25,549
And Kang uses this historical failure to highlight

231
00:10:25,549 --> 00:10:27,429
a critical tension in how we build technology

232
00:10:27,429 --> 00:10:30,830
today. He talks about the clash between disciplinary

233
00:10:30,830 --> 00:10:34,029
objectivity and mechanical objectivity. OK, let's

234
00:10:34,029 --> 00:10:36,590
break those down. What is disciplinary objectivity?

235
00:10:37,379 --> 00:10:39,820
Disciplinary objectivity is when we trust an

236
00:10:39,820 --> 00:10:42,159
expert's trained eye. We trust the doctor reading

237
00:10:42,159 --> 00:10:44,460
the x -ray, or we trust Kirsta because he says,

238
00:10:44,519 --> 00:10:46,899
hey, I have the discipline and the experience

239
00:10:46,899 --> 00:10:50,120
to see these hidden patterns. It relies completely

240
00:10:50,120 --> 00:10:53,559
on human judgment. Got it. And mechanical objectivity.

241
00:10:54,039 --> 00:10:57,259
Mechanical objectivity is when we trust a machine's

242
00:10:57,259 --> 00:10:59,980
rules. We build an algorithm to remove the human

243
00:10:59,980 --> 00:11:02,299
bias entirely. Let the machine do the measuring.

244
00:11:02,519 --> 00:11:04,840
So Kirste was basically trying to launder his

245
00:11:04,840 --> 00:11:07,340
disciplinary objectivity, his own gut feeling,

246
00:11:07,779 --> 00:11:09,600
through the mechanical objectivity of a fancy

247
00:11:09,600 --> 00:11:12,120
machine? That's exactly what he was doing. He

248
00:11:12,120 --> 00:11:14,759
claimed his method was a mechanical, infallible

249
00:11:14,759 --> 00:11:17,919
rule. But in reality, reading a spectrogram completely

250
00:11:17,919 --> 00:11:20,960
relied on his subjective, highly contested expertise.

251
00:11:21,200 --> 00:11:22,840
Because nobody else could do it. Right. When

252
00:11:22,840 --> 00:11:24,980
the wider scientific community tried to replicate

253
00:11:24,980 --> 00:11:28,279
his 99 % accuracy without him in the room guiding

254
00:11:28,279 --> 00:11:31,139
them, they couldn't do it. The scientific consensus

255
00:11:31,139 --> 00:11:33,639
wasn't just weak, it was literally non -existent.

256
00:11:33,960 --> 00:11:36,179
So if I'm understanding this correctly, Lawrence

257
00:11:36,179 --> 00:11:38,340
Kirste wasn't actually building a mechanical

258
00:11:38,340 --> 00:11:41,580
rule, he was building a machine that just outsourced

259
00:11:41,580 --> 00:11:43,639
the final judgment right back to his own eyeballs.

260
00:11:44,120 --> 00:11:46,419
And exactly. And because he was the only one

261
00:11:46,419 --> 00:11:49,620
who quote unquote knew how to read it, he became

262
00:11:49,620 --> 00:11:53,120
the bottleneck. I mean the 1979 report explicitly

263
00:11:53,120 --> 00:11:55,759
stated that voice identification was closer to

264
00:11:55,759 --> 00:11:58,460
an empirical art than a fully developed science.

265
00:11:58,649 --> 00:12:01,330
Yes, and they recommended a complete halt to

266
00:12:01,330 --> 00:12:03,950
forensic use until someone could establish a

267
00:12:03,950 --> 00:12:06,309
basic underlying scientific understanding of

268
00:12:06,309 --> 00:12:09,549
the human voice. But, and this is where Kang's

269
00:12:09,549 --> 00:12:13,129
book takes a pretty wild turn, the industry didn't

270
00:12:13,129 --> 00:12:15,870
stop. They didn't retreat to the biology labs

271
00:12:15,870 --> 00:12:18,669
to map the vocal tract. No, they didn't. They

272
00:12:18,669 --> 00:12:21,690
just ignored the biologists entirely. They bypassed

273
00:12:21,690 --> 00:12:24,700
them completely. Yes. They hit a scientific roadblock,

274
00:12:24,840 --> 00:12:27,460
so they just reframed the entire problem as an

275
00:12:27,460 --> 00:12:29,960
engineering and statistical challenge. If they

276
00:12:29,960 --> 00:12:32,179
couldn't understand the biology of why voices

277
00:12:32,179 --> 00:12:35,259
correlate to identity, they decided they just

278
00:12:35,259 --> 00:12:38,179
didn't need to. Right. They shifted to deep learning

279
00:12:38,179 --> 00:12:40,700
and what we now call automatic speaker recognition,

280
00:12:40,960 --> 00:12:43,700
or ASR. Okay, hold on. Break that down for me.

281
00:12:43,820 --> 00:12:46,139
How do you go from failing to map the biology

282
00:12:46,139 --> 00:12:49,080
to just solving it with math? I mean, they don't

283
00:12:49,080 --> 00:12:51,039
know the rules of the voice. How do they build

284
00:12:51,039 --> 00:12:53,549
the system? They did it by changing the paradigm

285
00:12:53,549 --> 00:12:56,970
from causation to correlation. The new logic

286
00:12:56,970 --> 00:12:59,350
became, look, we don't need a grand scientific

287
00:12:59,350 --> 00:13:02,330
theory if we can just feed a computer a massive

288
00:13:02,330 --> 00:13:05,230
amount of data and let it find statistical patterns

289
00:13:05,230 --> 00:13:07,450
on its own. They replaced the need for causal

290
00:13:07,450 --> 00:13:10,250
scientific theory with sheer brute force scale.

291
00:13:10,470 --> 00:13:13,250
Which leads us to these massive data sets. And

292
00:13:13,250 --> 00:13:15,870
Kang talks about Vox Celeb. I read that section

293
00:13:15,870 --> 00:13:18,529
and I was just stunned. Because if you want to

294
00:13:18,529 --> 00:13:20,990
train a highly precise audio recognition system,

295
00:13:21,210 --> 00:13:24,250
You'd think you would bring people into soundproof

296
00:13:24,250 --> 00:13:26,830
anechoic chambers, right? You'd think so. Use

297
00:13:26,830 --> 00:13:29,970
thousand -dollar microphones, get perfect, pristine

298
00:13:29,970 --> 00:13:32,190
audio samples. That's what Kirsta would have

299
00:13:32,190 --> 00:13:34,110
wanted. But that's the old scientific method.

300
00:13:34,450 --> 00:13:37,070
That's the old scientific method. But Vox Celeb

301
00:13:37,070 --> 00:13:39,370
didn't do that at all. They just wrote a script

302
00:13:39,370 --> 00:13:42,070
to scrape millions of utterances from YouTube

303
00:13:42,070 --> 00:13:45,129
celebrity interview videos. Yep. Red carpet interviews,

304
00:13:45,330 --> 00:13:46,970
late night talk shows, whatever they could find.

305
00:13:47,159 --> 00:13:49,379
It's full of background noise, people laughing,

306
00:13:49,620 --> 00:13:51,240
traffic, people talking over each other. It's

307
00:13:51,240 --> 00:13:53,799
an absolute mess. And that mess is exactly what

308
00:13:53,799 --> 00:13:56,580
makes modern neural networks so powerful. Deep

309
00:13:56,580 --> 00:13:59,000
learning models actually want that noise. They

310
00:13:59,000 --> 00:14:02,139
treat it as valuable variations. Really? Why?

311
00:14:02,879 --> 00:14:05,080
Because in machine learning, if you train an

312
00:14:05,080 --> 00:14:08,320
AI on perfect, pristine studio audio, it becomes

313
00:14:08,320 --> 00:14:12,100
what they call overfit. Overfit. OK, what does

314
00:14:12,100 --> 00:14:15,059
that mean? It means the second it hears a voice

315
00:14:15,059 --> 00:14:17,379
on a crappy cell phone connection in the real

316
00:14:17,379 --> 00:14:19,759
world, it fails. It doesn't know how to handle

317
00:14:19,759 --> 00:14:22,879
the imperfections. But by feeding it the chaotic,

318
00:14:23,379 --> 00:14:26,000
noisy YouTube data, the neural network learns

319
00:14:26,000 --> 00:14:28,980
to mathematically separate the underlying voice

320
00:14:28,980 --> 00:14:32,179
from the noise. It finds a pattern that actually

321
00:14:32,179 --> 00:14:36,840
survives the chaos. Oh, wow. Okay, so the machine

322
00:14:36,840 --> 00:14:39,080
is finding the pattern itself. Does that mean

323
00:14:39,080 --> 00:14:41,000
we finally achieve mechanical objectivity? Did

324
00:14:41,000 --> 00:14:43,320
we finally get the subjective human bias out

325
00:14:43,320 --> 00:14:47,080
of the loop? No. And this is one of Kang's darkest

326
00:14:47,080 --> 00:14:49,299
ironies in the entire book. We didn't solve the

327
00:14:49,299 --> 00:14:52,340
problem of human subjectivity at all. We just

328
00:14:52,340 --> 00:14:55,580
buried it deeper under layers of math. How so?

329
00:14:56,059 --> 00:14:58,799
Well, Kang compares Kirsta's highly subjective

330
00:14:58,799 --> 00:15:01,480
two -week training program in the 60s to the

331
00:15:01,480 --> 00:15:04,360
modern era of ASR. Any point specifically to

332
00:15:04,360 --> 00:15:07,539
a modern system called BatFox. Right. Batvox

333
00:15:07,539 --> 00:15:09,919
is this massive, widely used ASR system today.

334
00:15:10,059 --> 00:15:12,539
It's used by Unipol. It's used in forensic labs

335
00:15:12,539 --> 00:15:16,799
across 35 different countries. Exactly. And Agnesio,

336
00:15:16,919 --> 00:15:19,639
the company that created Batvox, has this tiered

337
00:15:19,639 --> 00:15:22,899
certification program for its users, basic, advanced,

338
00:15:23,179 --> 00:15:26,159
and expert levels. To become an expert, it requires

339
00:15:26,159 --> 00:15:28,840
three years of intensive contact with the software.

340
00:15:29,220 --> 00:15:32,240
Wait, I have to push back, because if Batbox

341
00:15:32,240 --> 00:15:34,720
requires three years of intensive training, doesn't

342
00:15:34,720 --> 00:15:37,000
that just mean it's a highly sophisticated, powerful

343
00:15:37,000 --> 00:15:40,279
tool that just requires a deeply skilled operator?

344
00:15:40,460 --> 00:15:43,659
Like a pilot flying a 747. Yeah, exactly. Why

345
00:15:43,659 --> 00:15:46,379
is Kang calling this a failure of objectivity

346
00:15:46,379 --> 00:15:49,799
rather than just recognizing it as deep expertise?

347
00:15:50,039 --> 00:15:52,580
It's a very fair question, and it's exactly the

348
00:15:52,580 --> 00:15:54,899
defense these companies use. But the difference

349
00:15:54,899 --> 00:15:57,559
lies in reproducibility and the black box nature

350
00:15:57,559 --> 00:16:00,470
of the tool itself. OK, unpack that. A pilot

351
00:16:00,470 --> 00:16:03,549
understands the aerodynamics of the 747. They

352
00:16:03,549 --> 00:16:07,190
know why it flies. But Kang notes that in 2017,

353
00:16:07,529 --> 00:16:10,789
reports showed only 20 % to 25 % of Batbox users

354
00:16:10,789 --> 00:16:13,149
had actually completed the certification. Oh,

355
00:16:13,149 --> 00:16:16,009
wow. Yeah. And more importantly, the system relies

356
00:16:16,009 --> 00:16:18,730
on statistical probability models that the human

357
00:16:18,730 --> 00:16:21,309
operator cannot manually verify. It's a black

358
00:16:21,309 --> 00:16:24,629
box. Aha. I see. So the software spits out a

359
00:16:24,629 --> 00:16:27,629
likelihood ratio saying, hey, there is a 95 %

360
00:16:27,629 --> 00:16:30,370
chance these two voices match. But the human

361
00:16:30,370 --> 00:16:32,450
sitting at the desk can't look under the hood

362
00:16:32,450 --> 00:16:35,899
and trace the exact math to see why it thinks

363
00:16:35,899 --> 00:16:38,639
that. Exactly. The human operator is still making

364
00:16:38,639 --> 00:16:40,759
a subjective judgment call at the end of the

365
00:16:40,759 --> 00:16:43,179
day. But now they are doing it based on the opaque

366
00:16:43,179 --> 00:16:45,559
output of a statistical software program they

367
00:16:45,559 --> 00:16:48,429
likely don't fully comprehend. So we just traded

368
00:16:48,429 --> 00:16:51,110
one flawed human expert in the 1960s reading

369
00:16:51,110 --> 00:16:54,389
a squiggly line on paper for another human in

370
00:16:54,389 --> 00:16:57,409
a forensic lab today clicking buttons on a screen.

371
00:16:58,090 --> 00:17:00,970
Precisely. The empirical art remains. It's just

372
00:17:00,970 --> 00:17:03,610
wearing a lab coat made of algorithms now. That

373
00:17:03,610 --> 00:17:07,009
is deeply unsettling. Because this historical

374
00:17:07,009 --> 00:17:09,329
shift, the moment the industry realized they

375
00:17:09,329 --> 00:17:10,990
didn't need to understand the science of the

376
00:17:10,990 --> 00:17:13,769
voice, but could just let computers find statistical

377
00:17:13,769 --> 00:17:17,000
patterns and massive data sets, That is the exact

378
00:17:17,000 --> 00:17:19,420
foundation that allows tech startups today to

379
00:17:19,420 --> 00:17:22,019
believe they can use your voice to predict practically

380
00:17:22,019 --> 00:17:25,660
anything. Yes. If the FBI can use math to guess

381
00:17:25,660 --> 00:17:29,299
identity, Silicon Valley wants to use math to

382
00:17:29,299 --> 00:17:31,720
guess your soul. Which brings us perfectly into

383
00:17:31,720 --> 00:17:33,740
the mechanics of how they actually try to pull

384
00:17:33,740 --> 00:17:36,339
that off because we need to talk about a concept

385
00:17:36,339 --> 00:17:40,359
Kang calls ground truth tracings or GTT. Okay.

386
00:17:40,630 --> 00:17:43,910
This is arguably the most vital methodological

387
00:17:43,910 --> 00:17:46,789
tool Kang introduces in the entire book. Let's

388
00:17:46,789 --> 00:17:50,130
get into it. What exactly is a ground truth tracing?

389
00:17:50,950 --> 00:17:53,470
So, GTT is essentially a framework for tracking

390
00:17:53,470 --> 00:17:57,289
how human engineers take incredibly messy, ambiguous

391
00:17:57,289 --> 00:18:00,049
real -world problems and literally force them

392
00:18:00,049 --> 00:18:03,130
into neat machine -learnable mathematical formats.

393
00:18:03,609 --> 00:18:06,029
Because an AI doesn't just learn abstract concepts

394
00:18:06,029 --> 00:18:08,250
on its own. Right. You can't just tell an AI,

395
00:18:08,569 --> 00:18:10,390
find a good employee. It doesn't know what good

396
00:18:10,390 --> 00:18:12,170
means. It doesn't even know what an employee

397
00:18:12,170 --> 00:18:14,349
is. It only knows numbers. Right. It's just a

398
00:18:14,349 --> 00:18:17,970
calculator on steroids. Exactly. So it has to

399
00:18:17,970 --> 00:18:20,109
be told exactly what it is looking for. It needs

400
00:18:20,109 --> 00:18:23,730
a ground truth. A ground truth. Yes. A reference

401
00:18:23,730 --> 00:18:26,210
point. A specific set of data that the human

402
00:18:26,210 --> 00:18:29,289
engineers have defined as 100 % factual and correct.

403
00:18:30,009 --> 00:18:32,450
The AI then uses that ground truth as the North

404
00:18:32,450 --> 00:18:35,099
Star to evaluate all new data coming in. And

405
00:18:35,099 --> 00:18:37,839
defining that ground truth requires so many massive

406
00:18:37,839 --> 00:18:40,259
human assumptions, compromises, and frankly,

407
00:18:40,420 --> 00:18:43,319
biases. Oh, incredible amounts of bias. Kang

408
00:18:43,319 --> 00:18:45,619
illustrates this with a chilling case study that

409
00:18:45,619 --> 00:18:48,440
I think everyone listening will find extremely

410
00:18:48,440 --> 00:18:51,819
relevant. It's about this voice AI startup that

411
00:18:51,819 --> 00:18:55,079
offers what is essentially a vocal resume. It

412
00:18:55,079 --> 00:18:58,539
is a prime example of GTT in action. So this

413
00:18:58,539 --> 00:19:01,220
startup claims its AI can analyze a job candidate's

414
00:19:01,220 --> 00:19:04,000
voice and To be clear, we are talking just the

415
00:19:04,000 --> 00:19:06,220
acoustic sound of it, not the actual words being

416
00:19:06,220 --> 00:19:09,019
spoken. And predict how well that candidate will

417
00:19:09,019 --> 00:19:11,619
fit into a corporate team. Right. It predicts

418
00:19:11,619 --> 00:19:14,380
their personality traits, their emotional stability,

419
00:19:14,380 --> 00:19:17,440
and how they will perform on the job. And the

420
00:19:17,440 --> 00:19:19,259
founder of this startup claims that technology

421
00:19:19,259 --> 00:19:21,920
is entirely independent of language or culture.

422
00:19:22,279 --> 00:19:25,720
Which sounds like literal magic. Or just pure

423
00:19:25,720 --> 00:19:28,420
science fiction. Because if you were just listening

424
00:19:28,420 --> 00:19:31,220
to the pitch and the cadence, How on earth do

425
00:19:31,220 --> 00:19:33,619
you train a machine to hear a concept as abstract

426
00:19:33,619 --> 00:19:36,859
as team fit? You do it through what is known

427
00:19:36,859 --> 00:19:39,960
as the example -based paradigm. And when Kang

428
00:19:39,960 --> 00:19:42,460
breaks down how they actually do it, it exposes

429
00:19:42,460 --> 00:19:47,000
a massive fundamental logical flaw. How so? Well,

430
00:19:47,519 --> 00:19:50,599
to train this AI, the startup doesn't rely on

431
00:19:50,599 --> 00:19:53,700
centuries of sociological, psychological, or

432
00:19:53,700 --> 00:19:55,599
organizational research about what makes a good

433
00:19:55,599 --> 00:19:58,099
team dynamic. They don't read the science, just

434
00:19:58,099 --> 00:20:00,710
like the FBI abandoned the biology. Exactly the

435
00:20:00,710 --> 00:20:02,849
same thing. Instead, they just go to the client

436
00:20:02,849 --> 00:20:04,890
company, let's say a big accounting firm, and

437
00:20:04,890 --> 00:20:07,569
they say, hey, give us voice recordings of your

438
00:20:07,569 --> 00:20:10,329
current good employees. The people you've already

439
00:20:10,329 --> 00:20:14,109
decided are top performers. Yes. OK, I want to

440
00:20:14,109 --> 00:20:16,089
slow down here because this is where the tech

441
00:20:16,089 --> 00:20:19,789
goes from sounding like magic to sounding incredibly

442
00:20:19,789 --> 00:20:22,569
mechanical. We really need to explain how a computer

443
00:20:22,569 --> 00:20:25,410
actually hears these good employees. Walk me

444
00:20:25,410 --> 00:20:27,740
through the mechanism. How do we get from a human

445
00:20:27,740 --> 00:20:30,319
speaking into a microphone to a machine deciding

446
00:20:30,319 --> 00:20:33,160
their personality? Okay, so it's all about extracting

447
00:20:33,160 --> 00:20:35,799
features and mapping them into a vector space.

448
00:20:36,160 --> 00:20:38,819
A vector space, okay. When you speak into a microphone,

449
00:20:39,019 --> 00:20:41,200
the sound wave is converted into digital data,

450
00:20:41,380 --> 00:20:44,259
zeros and ones. The AI software then combs through

451
00:20:44,259 --> 00:20:47,339
that digital audio and extracts acoustic features.

452
00:20:47,960 --> 00:20:50,660
What kind of features? Things like pitch, the

453
00:20:50,660 --> 00:20:53,440
variation in pitch, which they call jitter. The

454
00:20:53,440 --> 00:20:55,420
variation in loudness, which is called shimmer.

455
00:20:55,769 --> 00:20:58,250
the speed of speech, the length of the pauses

456
00:20:58,250 --> 00:21:00,910
between your words. There can be hundreds of

457
00:21:00,910 --> 00:21:03,609
these tiny microscopic acoustic measurements.

458
00:21:04,390 --> 00:21:06,829
So the AI completely strips away the words, strips

459
00:21:06,829 --> 00:21:09,230
away the meaning, and just creates a massive

460
00:21:09,230 --> 00:21:11,670
spreadsheet of these acoustic measurements for

461
00:21:11,670 --> 00:21:14,730
every single voice. Yes. And then it plots those

462
00:21:14,730 --> 00:21:17,450
measurements on a graph. OK. I'm with you. Now,

463
00:21:17,549 --> 00:21:19,609
if we were just measuring Two things, say, pitch

464
00:21:19,609 --> 00:21:22,630
and speed. We could plot it on a normal 2D piece

465
00:21:22,630 --> 00:21:24,990
of graph paper. Pitch is the y -axis. Speed is

466
00:21:24,990 --> 00:21:27,910
the x -axis. Your voice is a single dot on that

467
00:21:27,910 --> 00:21:30,990
paper. Simple enough. But the AI is measuring

468
00:21:30,990 --> 00:21:33,170
hundreds of features, right? So it doesn't plot

469
00:21:33,170 --> 00:21:36,410
them on a 2D graph. It plots them in a high -dimensional

470
00:21:36,410 --> 00:21:38,950
vector space. OK, wait. Let me slow you down

471
00:21:38,950 --> 00:21:41,430
there. What exactly does a high -dimensional

472
00:21:41,430 --> 00:21:44,150
vector space look like? Well, imagine a graph

473
00:21:44,150 --> 00:21:48,029
with 300 different axes all intersecting at once.

474
00:21:48,759 --> 00:21:52,369
I can't even picture that. A human brain literally

475
00:21:52,369 --> 00:21:54,730
cannot visualize it. It's impossible. But the

476
00:21:54,730 --> 00:21:57,250
computer can calculate it perfectly. It takes

477
00:21:57,250 --> 00:22:00,130
a complex, localized reality, a human voice with

478
00:22:00,130 --> 00:22:02,190
all its history, culture, and nuance, and it

479
00:22:02,190 --> 00:22:05,329
flattens it into a single mathematical coordinate

480
00:22:05,329 --> 00:22:08,549
in this massive, invisible, multi -dimensional

481
00:22:08,549 --> 00:22:12,589
space. Okay, I'm tracking with you now. So the

482
00:22:12,589 --> 00:22:15,490
AI takes all the good employees at this accounting

483
00:22:15,490 --> 00:22:18,650
firm, measure their voices, and plots them all

484
00:22:18,650 --> 00:22:22,599
in this 300 -dimensional space. Yes. And presumably

485
00:22:22,599 --> 00:22:24,700
because they all have similar communication styles

486
00:22:24,700 --> 00:22:28,019
that the managers obviously like, their dots

487
00:22:28,019 --> 00:22:30,660
cluster together in one specific area of the

488
00:22:30,660 --> 00:22:33,859
graph. Precisely. And that cluster becomes the

489
00:22:33,859 --> 00:22:36,700
ground truth. The AI draws a mathematical circle

490
00:22:36,700 --> 00:22:39,680
around that cluster and says, this specific area

491
00:22:39,680 --> 00:22:42,400
of acoustic coordinates equals a good employee.

492
00:22:42,589 --> 00:22:44,769
Here is where I have a massive problem with this.

493
00:22:45,309 --> 00:22:48,589
If this AI is just plotting points, and the ground

494
00:22:48,589 --> 00:22:51,390
truth is purely based on comparing my voice to

495
00:22:51,390 --> 00:22:52,769
the voices of the guys who already work at the

496
00:22:52,769 --> 00:22:55,190
firm, isn't this just a high -tech machine for

497
00:22:55,190 --> 00:22:57,690
enforcing corporate conformity? You've hit the

498
00:22:57,690 --> 00:22:59,609
exact critique, Kang, levels at this system.

499
00:23:00,150 --> 00:23:03,269
The AI isn't finding universal objective truths

500
00:23:03,269 --> 00:23:06,289
about human personality or competency. It is

501
00:23:06,289 --> 00:23:09,089
strictly mathematically measuring statistical

502
00:23:09,089 --> 00:23:11,829
proximity to the people who were already hired

503
00:23:11,829 --> 00:23:14,329
by the human managers of that specific company.

504
00:23:14,529 --> 00:23:16,529
It's just laundering human bias with the authority

505
00:23:16,529 --> 00:23:19,470
of machine learning. Exactly. I mean, let's say

506
00:23:19,470 --> 00:23:21,809
the current team at this firm is highly homogenous.

507
00:23:22,109 --> 00:23:25,809
Say they are predominantly white. male, neurotypical

508
00:23:25,809 --> 00:23:28,190
individuals from a similar socioeconomic class,

509
00:23:28,250 --> 00:23:30,609
and they all speak with a similar cadence and

510
00:23:30,609 --> 00:23:33,809
confidence level. Their specific vocal coordinates

511
00:23:33,809 --> 00:23:36,769
literally become the objective mathematical standard

512
00:23:36,769 --> 00:23:40,170
for a good employee. Yes. And the example -based

513
00:23:40,170 --> 00:23:42,849
paradigm assumes that because they have a large

514
00:23:42,849 --> 00:23:45,990
data set of employee voices, the underlying rules

515
00:23:45,990 --> 00:23:48,970
of good performance must be hidden in there somewhere.

516
00:23:48,970 --> 00:23:51,930
Right. They believe that data scale replaces

517
00:23:51,930 --> 00:23:54,980
the need for theory. But as Kang points out,

518
00:23:55,319 --> 00:23:58,299
vectorizing data inherently strips away the genesis

519
00:23:58,299 --> 00:24:00,900
and context of the voice. Right, because the

520
00:24:00,900 --> 00:24:03,619
AI doesn't know why you paused before you speak.

521
00:24:04,160 --> 00:24:07,299
No. It just knows that you paused. Like, you

522
00:24:07,299 --> 00:24:09,039
might be carefully translating a thought for

523
00:24:09,039 --> 00:24:11,559
your native language into English, which shows

524
00:24:11,559 --> 00:24:14,660
high cognitive ability and care. But the AI just

525
00:24:14,660 --> 00:24:17,430
measures quote unquote, long pause, compares

526
00:24:17,430 --> 00:24:19,410
it to the fast talking white guys in the cluster,

527
00:24:19,890 --> 00:24:22,549
and plots your dot far away from the good employee

528
00:24:22,549 --> 00:24:25,269
circle. Exactly. You're a judge not on who you

529
00:24:25,269 --> 00:24:27,890
are or your actual skills, but purely on your

530
00:24:27,890 --> 00:24:30,390
acoustic mathematical distance from the status

531
00:24:30,390 --> 00:24:33,650
quo. That's awful. If you have a African -American

532
00:24:33,650 --> 00:24:36,390
vernacular English cadence or a regional accent

533
00:24:36,390 --> 00:24:39,029
or a slight vocal tick, you are mathematically

534
00:24:39,029 --> 00:24:42,150
plotted further away. The AI automatically flags

535
00:24:42,150 --> 00:24:44,789
you as a bad fit. And because it's a machine

536
00:24:44,789 --> 00:24:47,529
outputting a number, the HR department assumes

537
00:24:47,529 --> 00:24:50,289
it's an objective, unbiased evaluation. Yes.

538
00:24:50,789 --> 00:24:54,170
They trust the math. This is deeply dystopian.

539
00:24:54,589 --> 00:24:57,049
We are buying software thinking it's an objective

540
00:24:57,049 --> 00:24:59,950
tool when it's actually just a digital mirror

541
00:24:59,950 --> 00:25:02,809
reflecting our existing biases back at us, but

542
00:25:02,809 --> 00:25:04,930
giving them this false weight of mathematical

543
00:25:04,930 --> 00:25:08,410
certainty. It's a huge problem. And this gets

544
00:25:08,410 --> 00:25:10,509
even more concerning when we move beyond predicting

545
00:25:10,509 --> 00:25:13,720
team fit, which is vague enough. to predicting

546
00:25:13,720 --> 00:25:16,279
something intensely subjective and fleeting.

547
00:25:16,759 --> 00:25:18,880
You're talking about human emotion. Exactly.

548
00:25:19,759 --> 00:25:21,960
If an AI can flatten personality into a math

549
00:25:21,960 --> 00:25:24,460
equation based purely on a cluster of existing

550
00:25:24,460 --> 00:25:27,519
employees, how does it handle emotion? Because

551
00:25:27,519 --> 00:25:30,440
emotion isn't static. It changes second by second.

552
00:25:30,799 --> 00:25:33,640
It's true. This is a massive subfield, and Kang

553
00:25:33,640 --> 00:25:36,039
dedicates a significant portion of his analysis

554
00:25:36,039 --> 00:25:38,579
to it. It's called speech emotion recognition,

555
00:25:38,900 --> 00:25:42,390
or SER. SER, right. It is an entire ecosystem

556
00:25:42,390 --> 00:25:45,410
of researchers and startups dedicated to teaching

557
00:25:45,410 --> 00:25:49,150
machines how to hear how we feel. And just like

558
00:25:49,150 --> 00:25:52,250
the vocal resume AI needed a ground truth, the

559
00:25:52,250 --> 00:25:55,619
SER field needs a ground truth for emotion. But

560
00:25:55,619 --> 00:25:58,119
before you can teach an AI to recognize an emotion,

561
00:25:58,279 --> 00:26:00,940
you run into a massive scientific roadblock.

562
00:26:01,619 --> 00:26:04,200
You have to actually define what an emotion is.

563
00:26:04,420 --> 00:26:06,819
And from what I understand, scientists don't

564
00:26:06,819 --> 00:26:09,930
even agree on that. It is one of the most fiercely

565
00:26:09,930 --> 00:26:12,230
debated topics in psychology and neuroscience

566
00:26:12,230 --> 00:26:15,009
today. There is no single agreed upon definition.

567
00:26:15,710 --> 00:26:18,109
Modern science, like the work of Lisa Feldman

568
00:26:18,109 --> 00:26:21,089
Barrett, increasingly views emotion as highly

569
00:26:21,089 --> 00:26:24,690
complex, highly contextual, and actually constructed

570
00:26:24,690 --> 00:26:28,210
in the moment by your brain based on past experience

571
00:26:28,210 --> 00:26:31,130
and current environment. Okay, so it isn't just

572
00:26:31,130 --> 00:26:34,079
a hardwired circuit. Right. So there isn't just

573
00:26:34,079 --> 00:26:36,819
one universal sadness circuit in the brain that

574
00:26:36,819 --> 00:26:38,940
lights up exactly the same way for every human

575
00:26:38,940 --> 00:26:41,779
on Earth. But that kind of messy, constructed,

576
00:26:42,140 --> 00:26:44,240
contextual reality is completely useless for

577
00:26:44,240 --> 00:26:46,700
machine learning. Completely useless. Machine

578
00:26:46,700 --> 00:26:49,500
learning needs neat. discrete labels. It needs

579
00:26:49,500 --> 00:26:52,960
buckets to put the data into. So the SER field

580
00:26:52,960 --> 00:26:55,960
largely ignores the modern contextual science

581
00:26:55,960 --> 00:26:58,339
of emotion and overwhelmingly relies on a much

582
00:26:58,339 --> 00:27:00,720
older, much more rigid framework called discrete

583
00:27:00,720 --> 00:27:03,039
emotion theory. This is Paul Ekman stuff, right?

584
00:27:03,839 --> 00:27:06,519
Yes. Ekman theorized that there are six basic

585
00:27:06,519 --> 00:27:09,380
biologically hardwired universal human emotions.

586
00:27:10,440 --> 00:27:14,039
Anger, disgust, fear, happiness, sadness, and

587
00:27:14,039 --> 00:27:16,559
surprise. I always think of this like the Pixar

588
00:27:16,559 --> 00:27:18,990
movie Inside Out. That's a perfect example. You

589
00:27:18,990 --> 00:27:21,289
just have these distinct little colored characters

590
00:27:21,289 --> 00:27:23,390
living in your head, pulling levers. It's a very

591
00:27:23,390 --> 00:27:25,630
clean, easy way to categorize feelings. It is.

592
00:27:25,930 --> 00:27:28,230
It's theoretically outdated, according to many

593
00:27:28,230 --> 00:27:30,609
modern psychologists, but it is mathematically

594
00:27:30,609 --> 00:27:33,269
incredibly convenient for AI developers. It gives

595
00:27:33,269 --> 00:27:36,650
them six neat little buckets. Right. But Kang

596
00:27:36,650 --> 00:27:39,569
exposes a much deeper problem here. The issue

597
00:27:39,569 --> 00:27:42,349
isn't just the six buckets. The issue is how

598
00:27:42,349 --> 00:27:44,470
the AI developers actually fill those buckets

599
00:27:44,470 --> 00:27:47,549
with data. How do you get the audio recordings

600
00:27:47,549 --> 00:27:49,849
to teach the machine what these six emotions

601
00:27:49,849 --> 00:27:52,569
sound like? And to explore this, Kang zeroes

602
00:27:52,569 --> 00:27:55,589
in on a specific benchmark dataset called the

603
00:27:55,589 --> 00:27:58,849
Montreal Effective Voices Dataset or MOV. Yes.

604
00:27:59,930 --> 00:28:02,779
This blew my mind. Because if you were building

605
00:28:02,779 --> 00:28:05,259
a data set to teach a machine what real human

606
00:28:05,259 --> 00:28:07,480
emotion sounds like, you would assume they went

607
00:28:07,480 --> 00:28:09,900
out and recorded people in real life highly emotional

608
00:28:09,900 --> 00:28:11,619
situations. You would think they'd use police

609
00:28:11,619 --> 00:28:14,440
body cams or emergency room audio or even just

610
00:28:14,440 --> 00:28:17,720
people reacting to surprising news, genuine spontaneous

611
00:28:17,720 --> 00:28:20,500
emotion. But they didn't do that at all. No.

612
00:28:21,099 --> 00:28:24,440
They brought actors into a studio. But, and this

613
00:28:24,440 --> 00:28:26,539
is key, they didn't even ask the actors to produce

614
00:28:26,539 --> 00:28:29,809
genuine emotional performances. What did they

615
00:28:29,809 --> 00:28:32,819
do? Kang explains that the MOV creators deliberately

616
00:28:32,819 --> 00:28:35,700
used a methodology called communication affect

617
00:28:35,700 --> 00:28:39,980
acting, or CEA. They explicitly instructed the

618
00:28:39,980 --> 00:28:42,980
actors to make stereotypical, highly exaggerated

619
00:28:42,980 --> 00:28:46,380
vocal bursts. Just sounds, no words? No words,

620
00:28:46,539 --> 00:28:49,359
just sounds. So if they wanted the fear bucket

621
00:28:49,359 --> 00:28:51,759
filled, they didn't want the actor to draw on

622
00:28:51,759 --> 00:28:54,519
a real memory of being terrified? No. Because

623
00:28:54,519 --> 00:28:57,240
that would be felt -experienced acting, or FEA,

624
00:28:57,599 --> 00:28:59,940
where the actor internalizes the emotion and

625
00:28:59,940 --> 00:29:03,079
the resulting sound is complex and subtle. The

626
00:29:03,079 --> 00:29:05,660
MOV creators didn't want subtle. They instructed

627
00:29:05,660 --> 00:29:07,859
the actors to strip away all personal context,

628
00:29:08,140 --> 00:29:10,839
all subtlety, and just deliver the most universally

629
00:29:10,839 --> 00:29:13,759
recognizable, almost cartoonishly pantomimed

630
00:29:13,759 --> 00:29:16,500
acoustic signal possible. I want to demonstrate

631
00:29:16,500 --> 00:29:19,400
this because it highlights how absurd the methodology

632
00:29:19,400 --> 00:29:22,039
is. Please do. If I use felt -experienced acting...

633
00:29:22,109 --> 00:29:24,170
And I imagine someone breaking into my house.

634
00:29:24,890 --> 00:29:27,329
My genuine fear response might be totally silent.

635
00:29:27,390 --> 00:29:30,150
I might just freeze. Or my breath might just

636
00:29:30,150 --> 00:29:32,670
get slightly shallow and I might whisper, who's

637
00:29:32,670 --> 00:29:35,609
there? Exactly. But the AI data set doesn't want

638
00:29:35,609 --> 00:29:38,329
that. It wants communication effect acting. It

639
00:29:38,329 --> 00:29:42,029
wants me to go, huh. Yes. That exaggerated gasp

640
00:29:42,029 --> 00:29:44,809
is exactly what gets recorded, vectorized, and

641
00:29:44,809 --> 00:29:47,750
plotted as the absolute mathematical ground truth

642
00:29:47,750 --> 00:29:51,640
for fear. That's insane. It is. And Kang makes

643
00:29:51,640 --> 00:29:54,319
a brilliant observation here. The creators of

644
00:29:54,319 --> 00:29:56,759
these data sets are not merely objective collectors

645
00:29:56,759 --> 00:30:00,240
of data. They are acting as emotional designers.

646
00:30:00,900 --> 00:30:03,359
Emotional designers. Yes. They are literally

647
00:30:03,359 --> 00:30:06,119
dictating to the machine what human emotion should

648
00:30:06,119 --> 00:30:08,160
sound like based on their own highly subjective

649
00:30:08,160 --> 00:30:10,720
assumptions of what a recognizable signal is.

650
00:30:11,099 --> 00:30:12,880
It's like a movie director telling an actor to

651
00:30:12,880 --> 00:30:16,109
act happy. in the most cliche, over -the -top,

652
00:30:16,369 --> 00:30:18,710
grinning, ear -to -ear way possible. Recording

653
00:30:18,710 --> 00:30:20,529
that audio and then telling a massive neural

654
00:30:20,529 --> 00:30:23,549
network, this is the only acceptable acoustic

655
00:30:23,549 --> 00:30:26,069
signature for happiness. And the real -world

656
00:30:26,069 --> 00:30:28,470
consequences of this design choice are immediate

657
00:30:28,470 --> 00:30:31,430
and severe. How so? Kang brings it back to the

658
00:30:31,430 --> 00:30:34,190
call centers. Companies are deploying this exact

659
00:30:34,190 --> 00:30:36,970
type of emotion AI to monitor customer service

660
00:30:36,970 --> 00:30:40,069
agents in real time, right now. So the AI is

661
00:30:40,069 --> 00:30:42,400
sitting on the line. listening to the agent's

662
00:30:42,400 --> 00:30:45,839
voice all day long, and scoring them on how happy

663
00:30:45,839 --> 00:30:49,259
or empathetic they sound, based entirely on its

664
00:30:49,259 --> 00:30:52,339
training from these exaggerated data sets. Yes.

665
00:30:52,599 --> 00:30:55,220
And in many cases, the worker's actual financial

666
00:30:55,220 --> 00:30:57,279
compensation, their bonuses, and their performance

667
00:30:57,279 --> 00:30:59,720
reviews are directly tied to these AI scores.

668
00:31:00,259 --> 00:31:02,839
So let me get this straight. If I'm a call center

669
00:31:02,839 --> 00:31:06,220
worker, and I am naturally a calm, soft -spoken

670
00:31:06,220 --> 00:31:08,920
person, I resolve the customer's issue politely

671
00:31:08,920 --> 00:31:12,240
and efficiently. Because I don't sound like a

672
00:31:12,240 --> 00:31:14,859
cartoon character, because my natural expression

673
00:31:14,859 --> 00:31:17,759
of contentment doesn't match the exaggerated

674
00:31:17,759 --> 00:31:20,440
communication affect acting birth that the AI

675
00:31:20,440 --> 00:31:24,599
was trained on, the AI flags me as unenthusiastic

676
00:31:24,599 --> 00:31:27,980
and it actually docks my pay? Precisely. The

677
00:31:27,980 --> 00:31:30,559
system demands an acoustic performance that matches

678
00:31:30,559 --> 00:31:34,039
its highly engineered ground truth. And Kang

679
00:31:34,039 --> 00:31:37,079
draws a vital, profound link to disability studies

680
00:31:37,079 --> 00:31:40,900
here. Oh, I can see where this is going. By training

681
00:31:40,900 --> 00:31:44,539
AI on these exaggerated neurotypical vocal bursts,

682
00:31:45,079 --> 00:31:47,920
the developers are establishing a rigid technological

683
00:31:47,920 --> 00:31:51,819
definition of normal. Right. It completely pathologizes

684
00:31:51,819 --> 00:31:53,980
anyone who falls outside that narrow acoustic

685
00:31:53,980 --> 00:31:57,559
window. Exactly. If you are neurodivergent and

686
00:31:57,559 --> 00:32:00,339
perhaps speak with a flatter affect, if you have

687
00:32:00,339 --> 00:32:02,839
a speech impediment, or even if your cultural

688
00:32:02,839 --> 00:32:05,119
background just expresses emotion with a different

689
00:32:05,119 --> 00:32:08,519
cadence, volume, or rhythm, Your voice is instantly

690
00:32:08,519 --> 00:32:11,420
and mathematically categorized as aberrant. Wow.

691
00:32:11,700 --> 00:32:14,299
You are flagged as angry when you are just passionate.

692
00:32:14,720 --> 00:32:16,880
You are flagged as sad when you are just intensely

693
00:32:16,880 --> 00:32:20,519
focused. The AI is actively policing human expression,

694
00:32:20,920 --> 00:32:23,200
forcing everyone to conform to a manufactured,

695
00:32:23,700 --> 00:32:26,799
engineered standard of emotion just to be recognized

696
00:32:26,799 --> 00:32:29,779
as competent. It's incredibly dystopian. Yeah.

697
00:32:29,880 --> 00:32:32,339
We are literally altering our own bodies, changing

698
00:32:32,339 --> 00:32:35,319
the way we breathe and speak in our daily jobs

699
00:32:35,319 --> 00:32:38,019
just to appease a machine that was trained on

700
00:32:38,019 --> 00:32:41,759
bad acting. It's a bizarre inversion of reality.

701
00:32:41,940 --> 00:32:45,420
It really is. But Kang doesn't stop there. We've

702
00:32:45,420 --> 00:32:47,680
spent all this time talking about how machines

703
00:32:47,680 --> 00:32:50,240
are taught this highly artificial way of listening

704
00:32:50,240 --> 00:32:53,299
to us. But the script completely flips when these

705
00:32:53,299 --> 00:32:56,559
exact same artificial mechanics are used to make

706
00:32:56,559 --> 00:33:00,250
machines speak. That transitions us beautifully

707
00:33:00,250 --> 00:33:03,750
into interlude 2 and chapter 4 of the book, which

708
00:33:03,750 --> 00:33:07,049
moves from listening machines to AI voice engines.

709
00:33:07,609 --> 00:33:09,329
We are entering the realm of text to speech,

710
00:33:09,690 --> 00:33:12,289
synthetic voices, and deep fake voice cloning.

711
00:33:12,849 --> 00:33:15,549
To ground this section, Kang brings in a really

712
00:33:15,549 --> 00:33:18,009
fascinating concept from film theory, going back

713
00:33:18,009 --> 00:33:20,730
to Michelle Sheehan again. It's the idea of the

714
00:33:20,730 --> 00:33:24,529
acousmatic voice. Yes. The acousmatic voice refers

715
00:33:24,529 --> 00:33:26,529
to a voice that you can hear, but you cannot

716
00:33:26,529 --> 00:33:29,299
see the body that is producing it. Think of the

717
00:33:29,299 --> 00:33:31,480
classic example, the booming wizard behind the

718
00:33:31,480 --> 00:33:34,960
curtain in The Wizard of Oz, or Howl 9000 in

719
00:33:34,960 --> 00:33:38,279
2001, A Space Odyssey. When we encounter an acousmatic

720
00:33:38,279 --> 00:33:41,000
voice, our human brains can't just leave it alone.

721
00:33:41,700 --> 00:33:44,200
We don't just hear the sound. We immediately

722
00:33:44,200 --> 00:33:48,380
try to imagine the source. Yes. Our brains automatically

723
00:33:48,380 --> 00:33:51,019
try to conjure a physical body to match the voice

724
00:33:51,019 --> 00:33:53,900
we are hearing. That mentally conjured body is

725
00:33:53,900 --> 00:33:56,539
what scholar Stephen Conner calls the vocalic

726
00:33:56,539 --> 00:34:00,559
body. the vocalic body. As humans, we inherently

727
00:34:00,559 --> 00:34:02,980
link voice and body. Our entire evolutionary

728
00:34:02,980 --> 00:34:05,359
history tells us that one cannot exist without

729
00:34:05,359 --> 00:34:07,299
the other. You can't have a voice without lungs,

730
00:34:07,500 --> 00:34:11,400
a throat, a mouth. But AI voice generation completely

731
00:34:11,400 --> 00:34:14,119
severs this biological link. Kong talks about

732
00:34:14,119 --> 00:34:17,360
the recent explosion in this technology as the

733
00:34:17,360 --> 00:34:21,619
chat GPT moment for audio. It really was. We

734
00:34:21,619 --> 00:34:24,320
have companies like Eleven Labs generating incredibly

735
00:34:24,320 --> 00:34:27,139
realistic human speech from text prompts. Sounds

736
00:34:27,139 --> 00:34:30,019
perfect. It has breath noises, it has pacing,

737
00:34:30,199 --> 00:34:33,920
it has micro hesitations. But to understand what

738
00:34:33,920 --> 00:34:36,659
is profoundly missing from these synthetic voices,

739
00:34:37,340 --> 00:34:39,420
Kong actually steps away from the tech developers

740
00:34:39,420 --> 00:34:42,219
and goes to interview professional human voice

741
00:34:42,219 --> 00:34:44,380
actors. And those interviews are some of the

742
00:34:44,380 --> 00:34:46,840
most illuminating parts of the book. He talks

743
00:34:46,840 --> 00:34:49,420
to actors named Jane and Christopher, and they

744
00:34:49,420 --> 00:34:52,639
explain that the human act of voicing is deeply,

745
00:34:53,199 --> 00:34:56,360
inherently, and inescapably embodied. It is a

746
00:34:56,360 --> 00:34:59,280
physical act. Right. Christopher talks about

747
00:34:59,280 --> 00:35:02,219
how when he voices a specific cartoon character,

748
00:35:02,719 --> 00:35:05,239
he doesn't just change his throat. He changes

749
00:35:05,239 --> 00:35:08,400
the entire physical geometry of his body. Exactly.

750
00:35:08,820 --> 00:35:11,480
He might square shoulders. thrust his chest out,

751
00:35:12,019 --> 00:35:13,760
alter the shape of his mouth, change the depth

752
00:35:13,760 --> 00:35:17,380
of his breathing. The voice is a byproduct of

753
00:35:17,380 --> 00:35:20,579
a physical posture. And Jane discusses the psychological

754
00:35:20,579 --> 00:35:23,360
aspect. Producing a specific emotional voice

755
00:35:23,360 --> 00:35:26,159
requires drawing on deep personal memories, lived

756
00:35:26,159 --> 00:35:29,260
experiences, and past traumas to find the right

757
00:35:29,260 --> 00:35:31,340
resonance and emotional truth in her delivery.

758
00:35:32,079 --> 00:35:34,920
Human voicing is a highly localized physical

759
00:35:34,920 --> 00:35:37,559
and psychological event. It comes from a specific

760
00:35:37,559 --> 00:35:40,639
body, in a specific place, with a specific history.

761
00:35:41,280 --> 00:35:43,719
But the machine act of synthesizing a voice has

762
00:35:43,719 --> 00:35:47,019
absolutely none of that. None. An AI voice engine

763
00:35:47,019 --> 00:35:49,440
doesn't have lungs, it doesn't have a diaphragm,

764
00:35:49,739 --> 00:35:51,760
it doesn't have a physical geometry to alter,

765
00:35:51,840 --> 00:35:54,019
and it certainly doesn't have memories to draw

766
00:35:54,019 --> 00:35:56,840
upon. Kang points out that an AI voice is created

767
00:35:56,840 --> 00:35:59,719
by statistically averaging hundreds, sometimes

768
00:35:59,719 --> 00:36:01,880
thousands, of individual human performances.

769
00:36:02,480 --> 00:36:05,030
So the neural network takes the breath intake

770
00:36:05,030 --> 00:36:07,690
of one actor, the gravely cadence of another,

771
00:36:07,829 --> 00:36:11,230
the emotional timbre of a third, and mathematically

772
00:36:11,230 --> 00:36:13,710
mashes them together in that high dimensional

773
00:36:13,710 --> 00:36:16,929
vector space we talked about earlier. Yep. It

774
00:36:16,929 --> 00:36:18,670
interpolates between the points to create a new

775
00:36:18,670 --> 00:36:22,610
average point. Exactly. It creates a manufactured

776
00:36:22,610 --> 00:36:25,309
surface effect. It sounds natural to our ears,

777
00:36:25,309 --> 00:36:27,670
but it is completely hollow. I was going to use

778
00:36:27,670 --> 00:36:30,949
the analogy that it's essentially a digital Frankenstein's

779
00:36:30,949 --> 00:36:33,670
monster. Right. It's stitched together from stolen

780
00:36:33,670 --> 00:36:36,230
human breaths and memories, and it walks and

781
00:36:36,230 --> 00:36:38,909
talks, but there is literally nobody home. Right.

782
00:36:39,840 --> 00:36:43,059
But actually, based on what Kang is saying, Frankenstein's

783
00:36:43,059 --> 00:36:45,840
monster is too generous an analogy. The monster

784
00:36:45,840 --> 00:36:48,940
actually had a physical body. The AI voice doesn't

785
00:36:48,940 --> 00:36:52,559
even have that. It is pure disembodied math masquerading

786
00:36:52,559 --> 00:36:54,820
as a human soul. That is a much more accurate

787
00:36:54,820 --> 00:36:58,139
way to look at it. It turns what was a unique

788
00:36:58,309 --> 00:37:01,570
embodied human expression into a completely disembodied

789
00:37:01,570 --> 00:37:03,730
mass -produced commodity. And once the voice

790
00:37:03,730 --> 00:37:06,210
is severed from the body and turned into a commodity,

791
00:37:06,670 --> 00:37:09,429
King argues, exploitation is the inevitable next

792
00:37:09,429 --> 00:37:12,230
step. And he backs that up with real -world examples.

793
00:37:12,949 --> 00:37:15,269
He highlights the Skyrim modding controversy.

794
00:37:15,869 --> 00:37:18,170
For anyone not familiar, Skyrim is a massively

795
00:37:18,170 --> 00:37:21,860
popular video game. Modders, fans who alter the

796
00:37:21,860 --> 00:37:24,940
game, started using AI tools to clone the voices

797
00:37:24,940 --> 00:37:27,579
of the professional human voice actors who originally

798
00:37:27,579 --> 00:37:30,000
worked on the game. And they did this without

799
00:37:30,000 --> 00:37:32,460
the actors' consent, and obviously without paying

800
00:37:32,460 --> 00:37:36,219
them. They took the actors' vocalic bodies, fed

801
00:37:36,219 --> 00:37:39,260
them into an AI engine, and forced these digital

802
00:37:39,260 --> 00:37:41,599
clones to say things the actors never agreed

803
00:37:41,599 --> 00:37:44,079
to say. Some of it was highly inappropriate or

804
00:37:44,079 --> 00:37:47,239
offensive material too. Yes. It's a profound

805
00:37:47,239 --> 00:37:49,860
violation of identity and a literal theft of

806
00:37:49,860 --> 00:37:52,500
labor. But the exploitation goes even deeper

807
00:37:52,500 --> 00:37:54,579
than just stealing an actor's performance for

808
00:37:54,579 --> 00:37:57,480
a video game. It extends into completely rewriting

809
00:37:57,480 --> 00:38:01,679
a person's cultural identity in real time. Kang

810
00:38:01,679 --> 00:38:04,699
brings up this terrifying phenomenon of AI passing.

811
00:38:05,260 --> 00:38:07,099
Remember the hook we started the show with, the

812
00:38:07,099 --> 00:38:09,239
call center in the Philippines? Yes, exactly.

813
00:38:09,780 --> 00:38:12,260
There are startups right now that are offering

814
00:38:12,260 --> 00:38:14,699
AI software to global call centers in places

815
00:38:14,699 --> 00:38:17,800
like the Philippines or India. The software acts

816
00:38:17,800 --> 00:38:21,659
as a real -time algorithmic filter. So how does

817
00:38:21,659 --> 00:38:24,400
it work in practice? The worker speaks in their

818
00:38:24,400 --> 00:38:27,019
natural voice with their natural accent and the

819
00:38:27,019 --> 00:38:30,159
AI intercepts the audio, strips away the worker's

820
00:38:30,159 --> 00:38:32,539
accent and replaces it on the customer's end

821
00:38:32,539 --> 00:38:35,900
with a synthesized white sounding American accent.

822
00:38:36,260 --> 00:38:38,909
And the company selling this software claim They

823
00:38:38,909 --> 00:38:41,510
are doing it for a noble reason, right? They

824
00:38:41,510 --> 00:38:44,250
say it reduces the racist abuse that offshore

825
00:38:44,250 --> 00:38:47,090
workers often face from angry American customers,

826
00:38:47,150 --> 00:38:50,030
which is a very real, horrific problem. It is

827
00:38:50,030 --> 00:38:52,309
a real problem, but the technological solution

828
00:38:52,309 --> 00:38:54,969
they are offering is profoundly disturbing. They

829
00:38:54,969 --> 00:38:58,170
are selling a digital, vocalic body. They are

830
00:38:58,170 --> 00:39:00,349
saying, your physical reality, your actual voice

831
00:39:00,349 --> 00:39:03,710
is the problem. It is unacceptable. The solution

832
00:39:03,710 --> 00:39:05,889
is to completely erase the worker's identity.

833
00:39:06,139 --> 00:39:08,719
substituting it with a mathematically generated

834
00:39:08,719 --> 00:39:11,940
shield of whiteness. It reinforces the exact

835
00:39:11,940 --> 00:39:14,820
racism it claims to be circumventing. It tells

836
00:39:14,820 --> 00:39:17,500
the customer, you're right. You only deserve

837
00:39:17,500 --> 00:39:20,559
to speak to someone who sounds like a white midwesterner.

838
00:39:20,900 --> 00:39:23,039
It treats the human voice like a cheap coat of

839
00:39:23,039 --> 00:39:25,599
paint that you can just digitally swap out to

840
00:39:25,599 --> 00:39:27,960
appease a biased algorithm or a biased customer.

841
00:39:28,699 --> 00:39:31,199
It is the ultimate flattening of the human experience.

842
00:39:31,699 --> 00:39:35,579
The unique localized voice is erased. And a disembodied

843
00:39:35,579 --> 00:39:38,039
statistical average of a standard accent is put

844
00:39:38,039 --> 00:39:41,360
in its place. OK, this brings up a massive logistical

845
00:39:41,360 --> 00:39:44,159
question for me. If these AI voices, whether

846
00:39:44,159 --> 00:39:47,159
it's the fake call center worker or the text

847
00:39:47,159 --> 00:39:50,019
to speech engine reading an audiobook, are just

848
00:39:50,019 --> 00:39:52,579
mechanical averages of stolen human performances,

849
00:39:53,199 --> 00:39:55,760
how does the tech industry actually evaluate

850
00:39:55,760 --> 00:39:58,599
them? What do you mean? Like, how do they prove

851
00:39:58,599 --> 00:40:01,400
to investors that their digital Frankenstein

852
00:40:01,400 --> 00:40:04,880
sounds good or natural? How do you measure the

853
00:40:04,880 --> 00:40:07,619
quality of a voice that doesn't actually exist?

854
00:40:07,900 --> 00:40:10,579
That is the perfect transition because it leads

855
00:40:10,579 --> 00:40:12,860
us into the final section of the book, Chapter

856
00:40:12,860 --> 00:40:15,579
Five. And this section reveals what might be

857
00:40:15,579 --> 00:40:18,760
the most bizarre, circular and revealing logic

858
00:40:18,760 --> 00:40:21,619
in the entire AI ecosystem. Oh, boy. It's all

859
00:40:21,619 --> 00:40:24,079
about how the industry measures progress. In

860
00:40:24,079 --> 00:40:26,639
machine learning, you absolutely need benchmark

861
00:40:26,639 --> 00:40:29,280
data sets. You need a standardized test to prove

862
00:40:29,280 --> 00:40:32,119
to the world that your new AI model is better

863
00:40:32,119 --> 00:40:34,559
than last year's model. Right. And for a long

864
00:40:34,559 --> 00:40:37,159
time, the benchmark for synthetic voices was

865
00:40:37,159 --> 00:40:40,639
pretty straightforward, right? It was human evaluation.

866
00:40:41,400 --> 00:40:44,000
They used a metric called the Mean Opinion Score,

867
00:40:44,260 --> 00:40:46,860
or MOS. Yes. They would generate a synthetic

868
00:40:46,860 --> 00:40:49,800
voice, play the audio clip for a room full of

869
00:40:49,800 --> 00:40:52,199
human listeners, and those humans would rate

870
00:40:52,199 --> 00:40:54,639
it on a scale of one to five for naturalness.

871
00:40:55,219 --> 00:40:58,380
One being terrible and robotic, five being completely

872
00:40:58,380 --> 00:41:00,869
indistinguishable from a real person. Exactly.

873
00:41:01,230 --> 00:41:03,309
The human ear was the ultimate arbiter of human

874
00:41:03,309 --> 00:41:06,690
speech. But there is a huge problem with the

875
00:41:06,690 --> 00:41:08,690
mean opinion score for the modern tech industry.

876
00:41:09,289 --> 00:41:11,269
Paying thousands of humans to sit in a room and

877
00:41:11,269 --> 00:41:13,969
listen to endless audio clips is incredibly slow.

878
00:41:14,150 --> 00:41:16,929
It's tedious. And most importantly, it is expensive.

879
00:41:17,389 --> 00:41:19,510
It creates a bottleneck in the development cycle.

880
00:41:19,789 --> 00:41:21,630
And Silicon Valley hates bottlenecks more than

881
00:41:21,630 --> 00:41:24,690
anything. So what do they do? In the name of

882
00:41:24,690 --> 00:41:26,969
speed and efficiency, the industry decided they

883
00:41:26,969 --> 00:41:29,610
needed to automate the evaluation process itself.

884
00:41:29,949 --> 00:41:32,610
Kang focuses on a specific dataset called the

885
00:41:32,610 --> 00:41:35,510
VoiceMOS benchmark. Researchers decided to take

886
00:41:35,510 --> 00:41:38,250
all the past human ratings, thousands of instances

887
00:41:38,250 --> 00:41:41,289
where humans scored an AI voice a 3 or a 4 or

888
00:41:41,289 --> 00:41:44,550
a 5, and they used that data to train an entirely

889
00:41:44,550 --> 00:41:47,630
new AI model. Okay, wait. Let me make sure I'm

890
00:41:47,630 --> 00:41:49,650
following the math here. The goal of this new

891
00:41:49,650 --> 00:41:52,309
AI model wasn't to generate speech. The goal

892
00:41:52,309 --> 00:41:55,130
of this new AI was to listen to a synthetic voice

893
00:41:55,130 --> 00:41:58,070
and, based on its training data, predict what

894
00:41:58,070 --> 00:42:00,519
rating a human would give it. if a human were

895
00:42:00,519 --> 00:42:02,420
actually listening. Wait, wait, wait. Hold on.

896
00:42:02,420 --> 00:42:04,619
This is incredible. Let me map out the full life

897
00:42:04,619 --> 00:42:08,329
cycle of this audio. Go for it. Step one, human

898
00:42:08,329 --> 00:42:10,409
actors perform to create the initial training

899
00:42:10,409 --> 00:42:15,409
data. Step two, AI model A statistically averages

900
00:42:15,409 --> 00:42:19,030
that data to synthesize a fake voice. Step three,

901
00:42:19,650 --> 00:42:21,690
instead of playing that fake voice for a human,

902
00:42:22,349 --> 00:42:25,610
AI model B is brought in to listen to AI model

903
00:42:25,610 --> 00:42:28,829
A until the engineers of model A sounds like

904
00:42:28,829 --> 00:42:31,349
a real human. That is exactly the architecture

905
00:42:31,349 --> 00:42:33,929
of the voice MOS system. That is like a student

906
00:42:33,929 --> 00:42:36,429
writing a test, grading their own test, and using

907
00:42:36,429 --> 00:42:38,489
an answer key that they hallucinated themselves.

908
00:42:39,030 --> 00:42:42,309
Pretty much. The human being, the only entity

909
00:42:42,309 --> 00:42:44,769
that actually possesses a voice and knows what

910
00:42:44,769 --> 00:42:47,789
one sounds like, has been entirely removed from

911
00:42:47,789 --> 00:42:51,789
the loop of evaluating what sounds human. It

912
00:42:51,789 --> 00:42:55,530
is a closed mechanical echo chamber. The AI ear

913
00:42:55,530 --> 00:42:59,269
is validating the AI voice. King uses this to

914
00:42:59,269 --> 00:43:02,710
expose the absolute absurdity of techno -utopianism.

915
00:43:03,230 --> 00:43:04,929
The machines are just talking to each other,

916
00:43:05,110 --> 00:43:07,769
optimizing for mathematical proximity to an old

917
00:43:07,769 --> 00:43:10,949
data set while completely detached from biological

918
00:43:10,949 --> 00:43:14,150
reality. It's a snake eating its own tail. Why

919
00:43:14,150 --> 00:43:16,510
would any rational engineer do this? It makes

920
00:43:16,510 --> 00:43:20,690
absolutely no logical scientific sense if your

921
00:43:20,690 --> 00:43:22,949
end goal is to create authentic human sounding

922
00:43:22,949 --> 00:43:25,690
speech that resonates with actual people. You're

923
00:43:25,690 --> 00:43:28,269
right. It makes no scientific sense at all, but

924
00:43:28,269 --> 00:43:30,909
it makes perfect unassailable economic sense.

925
00:43:31,210 --> 00:43:33,750
And this is where Kang ties the entire book together.

926
00:43:34,110 --> 00:43:36,269
He brings all of this back to the economics of

927
00:43:36,269 --> 00:43:39,170
the tech industry and specifically the mechanics

928
00:43:39,170 --> 00:43:42,889
of venture capital funding. Okay. In the absence

929
00:43:42,889 --> 00:43:46,320
of real grounded scientific benchmarks. Economic

930
00:43:46,320 --> 00:43:49,360
viability becomes the benchmark. Now the puzzle

931
00:43:49,360 --> 00:43:52,159
pieces are fitting together. Kang outlines what

932
00:43:52,159 --> 00:43:54,940
he calls the three P's that drive this entire

933
00:43:54,940 --> 00:43:58,599
ecosystem, progress, profit, and promise. The

934
00:43:58,599 --> 00:44:01,420
VC hype machine. Exactly. Think about companies

935
00:44:01,420 --> 00:44:04,679
like 11 Labs. Investors are pouring hundreds

936
00:44:04,679 --> 00:44:07,019
of millions, even billions of dollars into these

937
00:44:07,019 --> 00:44:10,309
voice AI startups. These massive valuations are

938
00:44:10,309 --> 00:44:12,610
not based on the scientific truth of the product.

939
00:44:12,869 --> 00:44:15,789
They aren't based on a deep fundamental understanding

940
00:44:15,789 --> 00:44:18,349
of human acoustics or biology. What are they

941
00:44:18,349 --> 00:44:21,309
based on? They are driven entirely by the promise

942
00:44:21,309 --> 00:44:24,210
of scale. The promise that you can mass produce

943
00:44:24,210 --> 00:44:27,469
the human voice as a cheap, infinitely reproducible

944
00:44:27,469 --> 00:44:30,110
commodity. And to keep that VC money flowing,

945
00:44:30,329 --> 00:44:32,170
you have to constantly show charts going up and

946
00:44:32,170 --> 00:44:34,710
to the right. You have to show progress. Yes.

947
00:44:34,940 --> 00:44:37,619
And when an AI can automatically evaluate another

948
00:44:37,619 --> 00:44:39,820
AI, you can run millions of tests in a single

949
00:44:39,820 --> 00:44:43,139
day. You can iterate and scale infinitely faster

950
00:44:43,139 --> 00:44:45,500
than if you had to wait for pesky slow human

951
00:44:45,500 --> 00:44:48,579
listeners to give you a mean opinion score. Wow.

952
00:44:48,920 --> 00:44:51,639
In the tech world, algorithmic speed and scale

953
00:44:51,639 --> 00:44:54,619
equal perceived progress. Perceived progress

954
00:44:54,619 --> 00:44:57,420
generates hype, hype generates VC funding, which

955
00:44:57,420 --> 00:45:00,619
equals profit. The financial imperative to scale

956
00:45:00,619 --> 00:45:03,320
quickly completely overrides the scientific reality

957
00:45:03,320 --> 00:45:06,099
of the technology. The machine evaluating the

958
00:45:06,099 --> 00:45:08,059
machine is just a mechanism to keep the money

959
00:45:08,059 --> 00:45:12,400
moving. So, after unpacking all of this, what

960
00:45:12,400 --> 00:45:15,659
does it all actually mean for us? We've traced

961
00:45:15,659 --> 00:45:18,340
this incredible, somewhat terrifying arc through

962
00:45:18,340 --> 00:45:22,250
Kang's book. We started in the 18th century with

963
00:45:22,250 --> 00:45:25,090
Jacques de Volcansant's mechanical duck, creating

964
00:45:25,090 --> 00:45:27,550
his surface -level illusion of biology to fool

965
00:45:27,550 --> 00:45:31,099
an audience. Right. Then we moved to the 1960s

966
00:45:31,099 --> 00:45:34,019
with the FBI and Lawrence Kirsta trying to turn

967
00:45:34,019 --> 00:45:37,480
messy voices into neat visual fingerprints and

968
00:45:37,480 --> 00:45:39,380
failing miserably because they ignored the environment

969
00:45:39,380 --> 00:45:42,179
and the equipment. And from that failure, we

970
00:45:42,179 --> 00:45:45,119
saw the pivot to deep learning, the total abandonment

971
00:45:45,119 --> 00:45:47,659
of causal science for statistical correlation,

972
00:45:48,420 --> 00:45:50,960
leading to systems like Batvox, where the human

973
00:45:50,960 --> 00:45:53,980
subjectivity is just buried under complex math.

974
00:45:54,139 --> 00:45:56,880
Right. And then we saw how that mathematical

975
00:45:56,880 --> 00:45:59,980
flattening is applied to everyday life. We have

976
00:45:59,980 --> 00:46:02,579
AI systems that are actively judging our personalities

977
00:46:02,579 --> 00:46:05,800
in our team fit and job interviews, purely based

978
00:46:05,800 --> 00:46:08,300
on our acoustic proximity to the people a company

979
00:46:08,300 --> 00:46:12,309
already hired. We have emotion AI policing our

980
00:46:12,309 --> 00:46:14,730
happiness in call centers, forcing people to

981
00:46:14,730 --> 00:46:17,329
sound like exaggerated cartoons or risk losing

982
00:46:17,329 --> 00:46:20,210
their pay. We saw the severing of the voice from

983
00:46:20,210 --> 00:46:22,949
the body entirely. cloning actors without permission

984
00:46:22,949 --> 00:46:25,809
and stripping accents in real time to create

985
00:46:25,809 --> 00:46:28,829
a mathematically generated shield of whiteness.

986
00:46:29,409 --> 00:46:31,670
And it culminates in the tech industry setting

987
00:46:31,670 --> 00:46:34,969
up systems to evaluate themselves, completely

988
00:46:34,969 --> 00:46:37,090
removing the human ear from the loop, all in

989
00:46:37,090 --> 00:46:40,489
the name of venture capital scale. Tying all

990
00:46:40,489 --> 00:46:43,230
of these disparate threads together is the overarching

991
00:46:43,230 --> 00:46:46,570
thesis of Kang's work. The book serves as a massive

992
00:46:46,570 --> 00:46:48,809
warning about the danger of using the computer

993
00:46:48,809 --> 00:46:52,030
as a metaphor for the human mind and the machine

994
00:46:52,030 --> 00:46:54,769
as a metaphor for the human body. Knowing the

995
00:46:54,769 --> 00:46:57,190
voice through artificial intelligence does not

996
00:46:57,190 --> 00:47:00,230
mean we have achieved a deeper, more profound

997
00:47:00,230 --> 00:47:02,849
understanding of humanity. Not at all. It means

998
00:47:02,849 --> 00:47:05,429
we have successfully rendered the voice mechanical.

999
00:47:06,110 --> 00:47:08,389
We have stripped away the physical body, the

1000
00:47:08,389 --> 00:47:11,590
cultural context, the genuine emotion and all

1001
00:47:11,590 --> 00:47:13,980
the beautiful messiness of human interaction,

1002
00:47:14,519 --> 00:47:17,260
just so we can force the voice into a high dimensional

1003
00:47:17,260 --> 00:47:20,920
vector space, turning it into something measurable,

1004
00:47:21,340 --> 00:47:24,460
predictable, and ultimately, sellable. We're

1005
00:47:24,460 --> 00:47:27,239
building these incredibly powerful, globally

1006
00:47:27,239 --> 00:47:30,000
deployed systems based on a highly flawed, deeply

1007
00:47:30,000 --> 00:47:33,159
flattened metaphor of what it actually means

1008
00:47:33,159 --> 00:47:36,739
to be a human being. And that brings us to a

1009
00:47:36,739 --> 00:47:38,539
final pretty chilling thought to leave you with

1010
00:47:38,539 --> 00:47:41,670
as we wrap up this deep dive. Yeah. If artificial

1011
00:47:41,670 --> 00:47:43,989
intelligence inherently relies on flattening

1012
00:47:43,989 --> 00:47:47,909
our messy, beautiful, physical, embodied voices

1013
00:47:47,909 --> 00:47:50,869
into neat mathematical averages in order to function,

1014
00:47:51,789 --> 00:47:53,630
then the real danger isn't that machines will

1015
00:47:53,630 --> 00:47:55,389
eventually figure out how to sound perfectly

1016
00:47:55,389 --> 00:47:58,849
human. No. The real danger is that to survive

1017
00:47:58,849 --> 00:48:01,849
in a world gated by these AI algorithms, whether

1018
00:48:01,849 --> 00:48:04,010
you're applying for a job and your voice is being

1019
00:48:04,010 --> 00:48:06,610
analyzed for fit, or you're calling customer

1020
00:48:06,610 --> 00:48:08,530
service and trying to get the bot to understand

1021
00:48:08,530 --> 00:48:10,909
your frustration, or you're just trying to be

1022
00:48:10,909 --> 00:48:12,989
recognized as competent in your everyday life,

1023
00:48:13,489 --> 00:48:16,369
we will have to start sounding more like machines.

1024
00:48:16,730 --> 00:48:19,030
To survive the system, we will have to adapt

1025
00:48:19,030 --> 00:48:21,809
our own messy footprints to fit perfectly into

1026
00:48:21,809 --> 00:48:25,679
their broken x -ray machines. Exactly. The diagnostic

1027
00:48:25,679 --> 00:48:28,000
muddy waters aren't clearing up anytime soon.

1028
00:48:28,519 --> 00:48:30,460
We're just learning how to drown in them quietly,

1029
00:48:31,059 --> 00:48:33,380
altering our own voices to appease the algorithm.

1030
00:48:34,239 --> 00:48:35,840
Keep that in mind the next time you ask your

1031
00:48:35,840 --> 00:48:38,380
phone for the weather, or leave a voicemail you

1032
00:48:38,380 --> 00:48:40,460
know is being transcribed by a server somewhere.

1033
00:48:41,260 --> 00:48:44,639
Whose voice are you really using? Yours? Or the

1034
00:48:44,639 --> 00:48:45,980
one the machine wants to hear?