EPISODE TRANSCRIPT
Machine Voices: A Sonic Study of Artificial Intelligence (Edward B. Kang)
Automatically transcribed from the episode audio. Some words and names may be inaccurate.
Listen to the episodeDownload textDownload captions
Right now, somewhere in a sprawling call center in the Philippines, there's a worker speaking into a headset, right? They're trying to help a customer troubleshoot a router or something. Right. But the voice coming out of the other end of that line, the voice the customer actually hears, it belongs to a white American from the Midwest who literally does not exist. It's completely fabricated. Exactly. The worker's natural accent, their cadence, the actual physical reality of their vocal tract, it's all being captured and then just stripped away. It's mathematically replaced in real time by this synthetic digital shield. Which is just a profound manipulation of human interaction when you really stop to think about it. And what's wild here is that the technology doing this isn't some distant science fiction concept. Now it's happening right now. Right. It is a highly funded, rapidly scaling reality. And it's fundamentally altering how we even define human speech. So today we are plunging into this terrifying but honestly totally fascinating world of machine voices. We're doing a massive deep dive into Edward B. Kang's meticulous and Frankly, paradigm -shifting book. It's called Machine Voices, a Sonic Study of Artificial Intelligence.
A really incredible piece of research. Oh, absolutely. So our mission today is to pull apart this text concept by concept for you. We are going to explore the hidden, often invisible, intersection between our human voices and the AI industry that is currently, well, devouring them. Devouring is the right word. And Kang's core premise here is incredibly compelling, because he starts with a contradiction. Like, I want you to just... Think about how we traditionally view the human voice. The French film theorist Michel Chian famously called the voice a strange object because it's intimate, right? It's elusive. You can't just hold it in your hand. Exactly. You can't touch it. It's deeply tied to our psychology, our physical bodies, our personal histories.
It's the thing that makes you you. I mean, if you call your mother from an unknown number and just say one single word, she instantly knows it's you. Yes. But Kang's book exposes how the AI industry looks at this strange object and sees a completely different puzzle. They definitely don't see the intimacy. Not at all. They don't want the voice to be elusive or mysterious. They are actively trying to transform it into something manageable, something measurable, and above all else, something sellable. Right. And to pull off that transformation, to turn this intimate human breath into a literal commodity, Kang argues the AI industry relies on three derivative rationales.
Yes. These are basically the foundational myths of the entire voice AI enterprise. Let's lay these out because they kind of dictate everything we're going to talk about today. What's the first one? So the first rationale is the belief that voices are knowable because they are mechanical. Meaning, if we just build the right machine, we can completely understand the human voice. OK. And the second? The second is kind of the inverse of that. It's that voices are inherently mechanical and therefore they are knowable. Like we aren't ghosts in a shell. We are just fleshy instruments producing acoustic data. Which naturally leads to the inevitable third step. Right. The ultimate conclusion. Because voices are both knowable and mechanical, they are replicable.
If it's just a mechanism, we can just build a digital version of that exact mechanism. OK, let me pause here and push back on that a bit, or at least bring this down to earth for the person listening right now. Sure. Because if you are listening to this on a smartphone, the microphone you just used an hour ago to ask for driving directions has already mapped the physical geometry of your throat. Oh, absolutely. When you say, hey Siri, or use voice to text, you might think you're just using a handy tool. But under the surface, your specific laughs, your pauses, that unique timbre you have, It's all being mined as raw data to feed those exact three rationales. That is the exact reality Kang is analyzing here. We are moving from a paradigm where the voice is this phenomenon of human experience, this messy biological reality, to a paradigm where the voice is literally just a system of mechanisms. But getting to this point where an AI can clone a voice or an algorithm claims it can tell if you're a good employee just by the sound of your pitch, that obviously didn't happen overnight. No, it took decades of trial and error.
Right. Kang traces this back to the very first incredibly clumsy attempts to visually capture the human voice. And that brings up this historical anecdote from the text that I think perfectly sets the stage for the modern AI obsession. You brought this up when we were prepping. Jacques de Volcanson's mechanical duck. Yes, the digesting duck. It's such a brilliant historical parallel that Kang uses. Set the scene for us. So we're talking about the 18th century. Vaucanson built this elaborate copper automaton duck. It could flap its wings, drink water, and supposedly, this was the big draw, it could eat grain, digest it, and then excrete it. And people were just mesmerized by this? Oh, completely. Okay. They thought he had mechanically replicated biology.
They thought he built a digestive tract. They thought it was a total fraud, right? Completely fake. The grain went into one hidden compartment, and this preloaded fake excrement was released from a... Totally different compartment. Wow. Yeah, the machine wasn't digesting anything at all It was just mimicking the surface effect of digestion. Okay, so how does Kang connect this copper duck to modern AI? He uses it to ask a really critical question when a computer mimics a human voice perfectly has it actually replicated human intelligence or human biology or Is it just another mechanical duck? You know, is it just producing a highly convincing surface effect without any of the underlying reality?
That is a phenomenal setup. Because the quest to map the underlying reality of the voice takes us straight into the FBI's historical obsession with something called voice prints. Yes, chapter one. Right. This is where the modern tech industry really started to grapple with mapping sound to identity. So we are in the 1960s and 70s, the FBI is desperate for new forensic tools. And there's this massive push to use speech spectrograms to identify criminals. Let's break down what a spectrogram actually is, just so we're all on the same page, because the mechanism here is crucial. Yeah, go for it. So, sound in the air is just changes in pressure, right? It's a wave. If you look at a standard audio file on your computer, you see a waveform. It just shows how loud the sound is over time. Right, it gets fatter when it's loud, skinnier when it's quiet, like a heart monitor. Exactly, like a heart monitor. But a waveform doesn't tell you what the sound actually is. A spectrogram is a much more complex visual map. It takes an audio signal and plots three things simultaneously.
Time is on the horizontal axis, frequency or pitch is on the vertical axis, and the amplitude, the loudness, is represented by how dark or brightly colored the specific marks are. So it's basically taking a slice of audio and running it through a prism, right? Breaking it down into all its individual frequencies so you can literally look at the physical shape of a word. That's a really great way to visualize it. And in the 60s, this self -proclaimed expert named Lawrence Kursta looked at these visual shapes and made a massive, totally audacious claim. What did he claim? He heavily promoted the idea that a visual spectrogram of a voice was just as unique and just as reliable as a human fingerprint. Hence the term voice print. I mean, it's great marketing. It was brilliant marketing. Through these experiments he designed himself, Kirsta claimed an astounding 99 % accuracy rate in identifying speakers just by looking at these visual charts. 99 %? Yeah. And he was so confident, he set up his own company, VoicePrint Laboratories, and began offering a two -week training program to certify law enforcement officers as experts. OK, I have to stop you there. Two weeks? He thought he could train a cop to look at a complex frequency chart for two weeks? and achieve 99 % accuracy in identifying a human being. Yes. With enough certainty to literally send them to prison. He did. And for a while, courts were actually accepting this evidence.
Juries were buying it. That's terrifying. It really is. But this is where the conflict between biology and machinery really explodes. Defense attorneys start pushing back hard. The scientific community started raising their eyebrows. It became a complete courtroom nightmare. So what happened? Well, in 1976, the FBI finally said, look, we need to know if this is real science. So they commissioned the National Academy of Sciences to investigate it. Right, the NAS. And they put together this heavy hitting committee. We're talking physicists, linguists, audiologists, lawyers. They spend years looking at Krista's methods. And in 1979, they release a massive report that essentially torches Krista's entire premise right down to the ground. And why they torched it is arguably the most important part of this whole story. Exactly. The committee pointed out a fundamental flaw in the metaphor of the voice print itself. Walk us through that flaw.
Well, think about a traditional fingerprint. It's a direct graphical representation of human anatomy. It is the actual physical ridges on your skin pressing onto a surface. It's a direct one -to -one impression. But a voiceogram isn't that at all. It's indirect. Yes. It is an indirect representation of an incredibly complex, multi -layered process. Because when you speak, the resulting sound isn't just a static imprint of your vocal cords. It's heavily affected by your physiological state. Like, do you have a cold?
It's affected by your psychological state. Are you stressed out during an interrogation? It's affected by the acoustic environment you're standing in. Is the room echoey? Is there traffic outside, right? Exactly. And crucially, it is profoundly altered by the mechanical equipment used to record it. A cheap mic sounds different than an expensive one. You simply cannot draw a straight objective line from a visual squiggle on a page to a person's unique biological identity. So treating a voice like a fingerprint is essentially like trying to read someone's soul through a footprint. But the footprint changes shape depending on if the person is tired or angry or carrying a heavy backpack or walking through mud. Right. And Kirsten was just looking at the mud and saying, yep, that's definitely Steve. That is a phenomenal analogy. The footprint is constantly shifting.
And Kang uses this historical failure to highlight a critical tension in how we build technology today. He talks about the clash between disciplinary objectivity and mechanical objectivity. OK, let's break those down. What is disciplinary objectivity? Disciplinary objectivity is when we trust an expert's trained eye. We trust the doctor reading the x -ray, or we trust Kirsta because he says, hey, I have the discipline and the experience to see these hidden patterns. It relies completely on human judgment. Got it. And mechanical objectivity.
Mechanical objectivity is when we trust a machine's rules. We build an algorithm to remove the human bias entirely. Let the machine do the measuring. So Kirste was basically trying to launder his disciplinary objectivity, his own gut feeling, through the mechanical objectivity of a fancy machine? That's exactly what he was doing. He claimed his method was a mechanical, infallible rule. But in reality, reading a spectrogram completely relied on his subjective, highly contested expertise. Because nobody else could do it. Right. When the wider scientific community tried to replicate his 99 % accuracy without him in the room guiding them, they couldn't do it. The scientific consensus wasn't just weak, it was literally non -existent.
So if I'm understanding this correctly, Lawrence Kirste wasn't actually building a mechanical rule, he was building a machine that just outsourced the final judgment right back to his own eyeballs. And exactly. And because he was the only one who quote unquote knew how to read it, he became the bottleneck. I mean the 1979 report explicitly stated that voice identification was closer to an empirical art than a fully developed science. Yes, and they recommended a complete halt to forensic use until someone could establish a basic underlying scientific understanding of the human voice. But, and this is where Kang's book takes a pretty wild turn, the industry didn't stop. They didn't retreat to the biology labs to map the vocal tract. No, they didn't. They just ignored the biologists entirely. They bypassed them completely. Yes. They hit a scientific roadblock, so they just reframed the entire problem as an engineering and statistical challenge. If they couldn't understand the biology of why voices correlate to identity, they decided they just didn't need to. Right. They shifted to deep learning and what we now call automatic speaker recognition, or ASR. Okay, hold on. Break that down for me.
How do you go from failing to map the biology to just solving it with math? I mean, they don't know the rules of the voice. How do they build the system? They did it by changing the paradigm from causation to correlation. The new logic became, look, we don't need a grand scientific theory if we can just feed a computer a massive amount of data and let it find statistical patterns on its own. They replaced the need for causal scientific theory with sheer brute force scale. Which leads us to these massive data sets. And Kang talks about Vox Celeb. I read that section and I was just stunned. Because if you want to train a highly precise audio recognition system, You'd think you would bring people into soundproof anechoic chambers, right? You'd think so. Use thousand -dollar microphones, get perfect, pristine audio samples. That's what Kirsta would have wanted. But that's the old scientific method.
That's the old scientific method. But Vox Celeb didn't do that at all. They just wrote a script to scrape millions of utterances from YouTube celebrity interview videos. Yep. Red carpet interviews, late night talk shows, whatever they could find. It's full of background noise, people laughing, traffic, people talking over each other. It's an absolute mess. And that mess is exactly what makes modern neural networks so powerful. Deep learning models actually want that noise. They treat it as valuable variations. Really? Why? Because in machine learning, if you train an AI on perfect, pristine studio audio, it becomes what they call overfit. Overfit. OK, what does that mean? It means the second it hears a voice on a crappy cell phone connection in the real world, it fails. It doesn't know how to handle the imperfections. But by feeding it the chaotic, noisy YouTube data, the neural network learns to mathematically separate the underlying voice from the noise. It finds a pattern that actually survives the chaos. Oh, wow. Okay, so the machine is finding the pattern itself. Does that mean we finally achieve mechanical objectivity? Did we finally get the subjective human bias out of the loop? No. And this is one of Kang's darkest ironies in the entire book. We didn't solve the problem of human subjectivity at all. We just buried it deeper under layers of math. How so?
Well, Kang compares Kirsta's highly subjective two -week training program in the 60s to the modern era of ASR. Any point specifically to a modern system called BatFox. Right. Batvox is this massive, widely used ASR system today. It's used by Unipol. It's used in forensic labs across 35 different countries. Exactly. And Agnesio, the company that created Batvox, has this tiered certification program for its users, basic, advanced, and expert levels. To become an expert, it requires three years of intensive contact with the software.
Wait, I have to push back, because if Batbox requires three years of intensive training, doesn't that just mean it's a highly sophisticated, powerful tool that just requires a deeply skilled operator? Like a pilot flying a 747. Yeah, exactly. Why is Kang calling this a failure of objectivity rather than just recognizing it as deep expertise? It's a very fair question, and it's exactly the defense these companies use. But the difference lies in reproducibility and the black box nature of the tool itself. OK, unpack that. A pilot understands the aerodynamics of the 747. They know why it flies. But Kang notes that in 2017, reports showed only 20 % to 25 % of Batbox users had actually completed the certification. Oh, wow. Yeah. And more importantly, the system relies on statistical probability models that the human operator cannot manually verify. It's a black box. Aha. I see. So the software spits out a likelihood ratio saying, hey, there is a 95 % chance these two voices match. But the human sitting at the desk can't look under the hood and trace the exact math to see why it thinks that. Exactly. The human operator is still making a subjective judgment call at the end of the day. But now they are doing it based on the opaque output of a statistical software program they likely don't fully comprehend. So we just traded one flawed human expert in the 1960s reading a squiggly line on paper for another human in a forensic lab today clicking buttons on a screen.
Precisely. The empirical art remains. It's just wearing a lab coat made of algorithms now. That is deeply unsettling. Because this historical shift, the moment the industry realized they didn't need to understand the science of the voice, but could just let computers find statistical patterns and massive data sets, That is the exact foundation that allows tech startups today to believe they can use your voice to predict practically anything. Yes. If the FBI can use math to guess identity, Silicon Valley wants to use math to guess your soul. Which brings us perfectly into the mechanics of how they actually try to pull that off because we need to talk about a concept Kang calls ground truth tracings or GTT. Okay.
This is arguably the most vital methodological tool Kang introduces in the entire book. Let's get into it. What exactly is a ground truth tracing? So, GTT is essentially a framework for tracking how human engineers take incredibly messy, ambiguous real -world problems and literally force them into neat machine -learnable mathematical formats. Because an AI doesn't just learn abstract concepts on its own. Right. You can't just tell an AI, find a good employee. It doesn't know what good means. It doesn't even know what an employee is. It only knows numbers. Right. It's just a calculator on steroids. Exactly. So it has to be told exactly what it is looking for. It needs a ground truth. A ground truth. Yes. A reference point. A specific set of data that the human engineers have defined as 100 % factual and correct.
The AI then uses that ground truth as the North Star to evaluate all new data coming in. And defining that ground truth requires so many massive human assumptions, compromises, and frankly, biases. Oh, incredible amounts of bias. Kang illustrates this with a chilling case study that I think everyone listening will find extremely relevant. It's about this voice AI startup that offers what is essentially a vocal resume. It is a prime example of GTT in action. So this startup claims its AI can analyze a job candidate's voice and To be clear, we are talking just the acoustic sound of it, not the actual words being spoken. And predict how well that candidate will fit into a corporate team. Right. It predicts their personality traits, their emotional stability, and how they will perform on the job. And the founder of this startup claims that technology is entirely independent of language or culture.
Which sounds like literal magic. Or just pure science fiction. Because if you were just listening to the pitch and the cadence, How on earth do you train a machine to hear a concept as abstract as team fit? You do it through what is known as the example -based paradigm. And when Kang breaks down how they actually do it, it exposes a massive fundamental logical flaw. How so? Well, to train this AI, the startup doesn't rely on centuries of sociological, psychological, or organizational research about what makes a good team dynamic. They don't read the science, just like the FBI abandoned the biology. Exactly the same thing. Instead, they just go to the client company, let's say a big accounting firm, and they say, hey, give us voice recordings of your current good employees. The people you've already decided are top performers. Yes. OK, I want to slow down here because this is where the tech goes from sounding like magic to sounding incredibly mechanical. We really need to explain how a computer actually hears these good employees. Walk me through the mechanism. How do we get from a human speaking into a microphone to a machine deciding their personality? Okay, so it's all about extracting features and mapping them into a vector space.
A vector space, okay. When you speak into a microphone, the sound wave is converted into digital data, zeros and ones. The AI software then combs through that digital audio and extracts acoustic features. What kind of features? Things like pitch, the variation in pitch, which they call jitter. The variation in loudness, which is called shimmer. the speed of speech, the length of the pauses between your words. There can be hundreds of these tiny microscopic acoustic measurements. So the AI completely strips away the words, strips away the meaning, and just creates a massive spreadsheet of these acoustic measurements for every single voice. Yes. And then it plots those measurements on a graph. OK. I'm with you. Now, if we were just measuring Two things, say, pitch and speed. We could plot it on a normal 2D piece of graph paper. Pitch is the y -axis. Speed is the x -axis. Your voice is a single dot on that paper. Simple enough. But the AI is measuring hundreds of features, right? So it doesn't plot them on a 2D graph. It plots them in a high -dimensional vector space. OK, wait. Let me slow you down there. What exactly does a high -dimensional vector space look like? Well, imagine a graph with 300 different axes all intersecting at once.
I can't even picture that. A human brain literally cannot visualize it. It's impossible. But the computer can calculate it perfectly. It takes a complex, localized reality, a human voice with all its history, culture, and nuance, and it flattens it into a single mathematical coordinate in this massive, invisible, multi -dimensional space. Okay, I'm tracking with you now. So the AI takes all the good employees at this accounting firm, measure their voices, and plots them all in this 300 -dimensional space. Yes. And presumably because they all have similar communication styles that the managers obviously like, their dots cluster together in one specific area of the graph. Precisely. And that cluster becomes the ground truth. The AI draws a mathematical circle around that cluster and says, this specific area of acoustic coordinates equals a good employee.
Here is where I have a massive problem with this. If this AI is just plotting points, and the ground truth is purely based on comparing my voice to the voices of the guys who already work at the firm, isn't this just a high -tech machine for enforcing corporate conformity? You've hit the exact critique, Kang, levels at this system. The AI isn't finding universal objective truths about human personality or competency. It is strictly mathematically measuring statistical proximity to the people who were already hired by the human managers of that specific company.
It's just laundering human bias with the authority of machine learning. Exactly. I mean, let's say the current team at this firm is highly homogenous. Say they are predominantly white. male, neurotypical individuals from a similar socioeconomic class, and they all speak with a similar cadence and confidence level. Their specific vocal coordinates literally become the objective mathematical standard for a good employee. Yes. And the example -based paradigm assumes that because they have a large data set of employee voices, the underlying rules of good performance must be hidden in there somewhere.
Right. They believe that data scale replaces the need for theory. But as Kang points out, vectorizing data inherently strips away the genesis and context of the voice. Right, because the AI doesn't know why you paused before you speak. No. It just knows that you paused. Like, you might be carefully translating a thought for your native language into English, which shows high cognitive ability and care. But the AI just measures quote unquote, long pause, compares it to the fast talking white guys in the cluster, and plots your dot far away from the good employee circle. Exactly. You're a judge not on who you are or your actual skills, but purely on your acoustic mathematical distance from the status quo. That's awful. If you have a African -American vernacular English cadence or a regional accent or a slight vocal tick, you are mathematically plotted further away. The AI automatically flags you as a bad fit. And because it's a machine outputting a number, the HR department assumes it's an objective, unbiased evaluation. Yes.
They trust the math. This is deeply dystopian. We are buying software thinking it's an objective tool when it's actually just a digital mirror reflecting our existing biases back at us, but giving them this false weight of mathematical certainty. It's a huge problem. And this gets even more concerning when we move beyond predicting team fit, which is vague enough. to predicting something intensely subjective and fleeting. You're talking about human emotion. Exactly. If an AI can flatten personality into a math equation based purely on a cluster of existing employees, how does it handle emotion? Because emotion isn't static. It changes second by second.
It's true. This is a massive subfield, and Kang dedicates a significant portion of his analysis to it. It's called speech emotion recognition, or SER. SER, right. It is an entire ecosystem of researchers and startups dedicated to teaching machines how to hear how we feel. And just like the vocal resume AI needed a ground truth, the SER field needs a ground truth for emotion. But before you can teach an AI to recognize an emotion, you run into a massive scientific roadblock.
You have to actually define what an emotion is. And from what I understand, scientists don't even agree on that. It is one of the most fiercely debated topics in psychology and neuroscience today. There is no single agreed upon definition. Modern science, like the work of Lisa Feldman Barrett, increasingly views emotion as highly complex, highly contextual, and actually constructed in the moment by your brain based on past experience and current environment. Okay, so it isn't just a hardwired circuit. Right. So there isn't just one universal sadness circuit in the brain that lights up exactly the same way for every human on Earth. But that kind of messy, constructed, contextual reality is completely useless for machine learning. Completely useless. Machine learning needs neat. discrete labels. It needs buckets to put the data into. So the SER field largely ignores the modern contextual science of emotion and overwhelmingly relies on a much older, much more rigid framework called discrete emotion theory. This is Paul Ekman stuff, right?
Yes. Ekman theorized that there are six basic biologically hardwired universal human emotions. Anger, disgust, fear, happiness, sadness, and surprise. I always think of this like the Pixar movie Inside Out. That's a perfect example. You just have these distinct little colored characters living in your head, pulling levers. It's a very clean, easy way to categorize feelings. It is. It's theoretically outdated, according to many modern psychologists, but it is mathematically incredibly convenient for AI developers. It gives them six neat little buckets. Right. But Kang exposes a much deeper problem here. The issue isn't just the six buckets. The issue is how the AI developers actually fill those buckets with data. How do you get the audio recordings to teach the machine what these six emotions sound like? And to explore this, Kang zeroes in on a specific benchmark dataset called the Montreal Effective Voices Dataset or MOV. Yes.
This blew my mind. Because if you were building a data set to teach a machine what real human emotion sounds like, you would assume they went out and recorded people in real life highly emotional situations. You would think they'd use police body cams or emergency room audio or even just people reacting to surprising news, genuine spontaneous emotion. But they didn't do that at all. No. They brought actors into a studio. But, and this is key, they didn't even ask the actors to produce genuine emotional performances. What did they do? Kang explains that the MOV creators deliberately used a methodology called communication affect acting, or CEA. They explicitly instructed the actors to make stereotypical, highly exaggerated vocal bursts. Just sounds, no words? No words, just sounds. So if they wanted the fear bucket filled, they didn't want the actor to draw on a real memory of being terrified? No. Because that would be felt -experienced acting, or FEA, where the actor internalizes the emotion and the resulting sound is complex and subtle. The MOV creators didn't want subtle. They instructed the actors to strip away all personal context, all subtlety, and just deliver the most universally recognizable, almost cartoonishly pantomimed acoustic signal possible. I want to demonstrate this because it highlights how absurd the methodology is. Please do. If I use felt -experienced acting...
And I imagine someone breaking into my house. My genuine fear response might be totally silent. I might just freeze. Or my breath might just get slightly shallow and I might whisper, who's there? Exactly. But the AI data set doesn't want that. It wants communication effect acting. It wants me to go, huh. Yes. That exaggerated gasp is exactly what gets recorded, vectorized, and plotted as the absolute mathematical ground truth for fear. That's insane. It is. And Kang makes a brilliant observation here. The creators of these data sets are not merely objective collectors of data. They are acting as emotional designers.
Emotional designers. Yes. They are literally dictating to the machine what human emotion should sound like based on their own highly subjective assumptions of what a recognizable signal is. It's like a movie director telling an actor to act happy. in the most cliche, over -the -top, grinning, ear -to -ear way possible. Recording that audio and then telling a massive neural network, this is the only acceptable acoustic signature for happiness. And the real -world consequences of this design choice are immediate and severe. How so? Kang brings it back to the call centers. Companies are deploying this exact type of emotion AI to monitor customer service agents in real time, right now. So the AI is sitting on the line. listening to the agent's voice all day long, and scoring them on how happy or empathetic they sound, based entirely on its training from these exaggerated data sets. Yes.
And in many cases, the worker's actual financial compensation, their bonuses, and their performance reviews are directly tied to these AI scores. So let me get this straight. If I'm a call center worker, and I am naturally a calm, soft -spoken person, I resolve the customer's issue politely and efficiently. Because I don't sound like a cartoon character, because my natural expression of contentment doesn't match the exaggerated communication affect acting birth that the AI was trained on, the AI flags me as unenthusiastic and it actually docks my pay? Precisely. The system demands an acoustic performance that matches its highly engineered ground truth. And Kang draws a vital, profound link to disability studies here. Oh, I can see where this is going. By training AI on these exaggerated neurotypical vocal bursts, the developers are establishing a rigid technological definition of normal. Right. It completely pathologizes anyone who falls outside that narrow acoustic window. Exactly. If you are neurodivergent and perhaps speak with a flatter affect, if you have a speech impediment, or even if your cultural background just expresses emotion with a different cadence, volume, or rhythm, Your voice is instantly and mathematically categorized as aberrant. Wow.
You are flagged as angry when you are just passionate. You are flagged as sad when you are just intensely focused. The AI is actively policing human expression, forcing everyone to conform to a manufactured, engineered standard of emotion just to be recognized as competent. It's incredibly dystopian. Yeah. We are literally altering our own bodies, changing the way we breathe and speak in our daily jobs just to appease a machine that was trained on bad acting. It's a bizarre inversion of reality.
It really is. But Kang doesn't stop there. We've spent all this time talking about how machines are taught this highly artificial way of listening to us. But the script completely flips when these exact same artificial mechanics are used to make machines speak. That transitions us beautifully into interlude 2 and chapter 4 of the book, which moves from listening machines to AI voice engines. We are entering the realm of text to speech, synthetic voices, and deep fake voice cloning.
To ground this section, Kang brings in a really fascinating concept from film theory, going back to Michelle Sheehan again. It's the idea of the acousmatic voice. Yes. The acousmatic voice refers to a voice that you can hear, but you cannot see the body that is producing it. Think of the classic example, the booming wizard behind the curtain in The Wizard of Oz, or Howl 9000 in 2001, A Space Odyssey. When we encounter an acousmatic voice, our human brains can't just leave it alone. We don't just hear the sound. We immediately try to imagine the source. Yes. Our brains automatically try to conjure a physical body to match the voice we are hearing. That mentally conjured body is what scholar Stephen Conner calls the vocalic body. the vocalic body. As humans, we inherently link voice and body. Our entire evolutionary history tells us that one cannot exist without the other. You can't have a voice without lungs, a throat, a mouth. But AI voice generation completely severs this biological link. Kong talks about the recent explosion in this technology as the chat GPT moment for audio. It really was. We have companies like Eleven Labs generating incredibly realistic human speech from text prompts. Sounds perfect. It has breath noises, it has pacing, it has micro hesitations. But to understand what is profoundly missing from these synthetic voices, Kong actually steps away from the tech developers and goes to interview professional human voice actors. And those interviews are some of the most illuminating parts of the book. He talks to actors named Jane and Christopher, and they explain that the human act of voicing is deeply, inherently, and inescapably embodied. It is a physical act. Right. Christopher talks about how when he voices a specific cartoon character, he doesn't just change his throat. He changes the entire physical geometry of his body. Exactly.
He might square shoulders. thrust his chest out, alter the shape of his mouth, change the depth of his breathing. The voice is a byproduct of a physical posture. And Jane discusses the psychological aspect. Producing a specific emotional voice requires drawing on deep personal memories, lived experiences, and past traumas to find the right resonance and emotional truth in her delivery. Human voicing is a highly localized physical and psychological event. It comes from a specific body, in a specific place, with a specific history.
But the machine act of synthesizing a voice has absolutely none of that. None. An AI voice engine doesn't have lungs, it doesn't have a diaphragm, it doesn't have a physical geometry to alter, and it certainly doesn't have memories to draw upon. Kang points out that an AI voice is created by statistically averaging hundreds, sometimes thousands, of individual human performances. So the neural network takes the breath intake of one actor, the gravely cadence of another, the emotional timbre of a third, and mathematically mashes them together in that high dimensional vector space we talked about earlier. Yep. It interpolates between the points to create a new average point. Exactly. It creates a manufactured surface effect. It sounds natural to our ears, but it is completely hollow. I was going to use the analogy that it's essentially a digital Frankenstein's monster. Right. It's stitched together from stolen human breaths and memories, and it walks and talks, but there is literally nobody home. Right.
But actually, based on what Kang is saying, Frankenstein's monster is too generous an analogy. The monster actually had a physical body. The AI voice doesn't even have that. It is pure disembodied math masquerading as a human soul. That is a much more accurate way to look at it. It turns what was a unique embodied human expression into a completely disembodied mass -produced commodity. And once the voice is severed from the body and turned into a commodity, King argues, exploitation is the inevitable next step. And he backs that up with real -world examples.
He highlights the Skyrim modding controversy. For anyone not familiar, Skyrim is a massively popular video game. Modders, fans who alter the game, started using AI tools to clone the voices of the professional human voice actors who originally worked on the game. And they did this without the actors' consent, and obviously without paying them. They took the actors' vocalic bodies, fed them into an AI engine, and forced these digital clones to say things the actors never agreed to say. Some of it was highly inappropriate or offensive material too. Yes. It's a profound violation of identity and a literal theft of labor. But the exploitation goes even deeper than just stealing an actor's performance for a video game. It extends into completely rewriting a person's cultural identity in real time. Kang brings up this terrifying phenomenon of AI passing.
Remember the hook we started the show with, the call center in the Philippines? Yes, exactly. There are startups right now that are offering AI software to global call centers in places like the Philippines or India. The software acts as a real -time algorithmic filter. So how does it work in practice? The worker speaks in their natural voice with their natural accent and the AI intercepts the audio, strips away the worker's accent and replaces it on the customer's end with a synthesized white sounding American accent.
And the company selling this software claim They are doing it for a noble reason, right? They say it reduces the racist abuse that offshore workers often face from angry American customers, which is a very real, horrific problem. It is a real problem, but the technological solution they are offering is profoundly disturbing. They are selling a digital, vocalic body. They are saying, your physical reality, your actual voice is the problem. It is unacceptable. The solution is to completely erase the worker's identity. substituting it with a mathematically generated shield of whiteness. It reinforces the exact racism it claims to be circumventing. It tells the customer, you're right. You only deserve to speak to someone who sounds like a white midwesterner.
It treats the human voice like a cheap coat of paint that you can just digitally swap out to appease a biased algorithm or a biased customer. It is the ultimate flattening of the human experience. The unique localized voice is erased. And a disembodied statistical average of a standard accent is put in its place. OK, this brings up a massive logistical question for me. If these AI voices, whether it's the fake call center worker or the text to speech engine reading an audiobook, are just mechanical averages of stolen human performances, how does the tech industry actually evaluate them? What do you mean? Like, how do they prove to investors that their digital Frankenstein sounds good or natural? How do you measure the quality of a voice that doesn't actually exist?
That is the perfect transition because it leads us into the final section of the book, Chapter Five. And this section reveals what might be the most bizarre, circular and revealing logic in the entire AI ecosystem. Oh, boy. It's all about how the industry measures progress. In machine learning, you absolutely need benchmark data sets. You need a standardized test to prove to the world that your new AI model is better than last year's model. Right. And for a long time, the benchmark for synthetic voices was pretty straightforward, right? It was human evaluation.
They used a metric called the Mean Opinion Score, or MOS. Yes. They would generate a synthetic voice, play the audio clip for a room full of human listeners, and those humans would rate it on a scale of one to five for naturalness. One being terrible and robotic, five being completely indistinguishable from a real person. Exactly. The human ear was the ultimate arbiter of human speech. But there is a huge problem with the mean opinion score for the modern tech industry. Paying thousands of humans to sit in a room and listen to endless audio clips is incredibly slow.
It's tedious. And most importantly, it is expensive. It creates a bottleneck in the development cycle. And Silicon Valley hates bottlenecks more than anything. So what do they do? In the name of speed and efficiency, the industry decided they needed to automate the evaluation process itself. Kang focuses on a specific dataset called the VoiceMOS benchmark. Researchers decided to take all the past human ratings, thousands of instances where humans scored an AI voice a 3 or a 4 or a 5, and they used that data to train an entirely new AI model. Okay, wait. Let me make sure I'm following the math here. The goal of this new AI model wasn't to generate speech. The goal of this new AI was to listen to a synthetic voice and, based on its training data, predict what rating a human would give it. if a human were actually listening. Wait, wait, wait. Hold on.
This is incredible. Let me map out the full life cycle of this audio. Go for it. Step one, human actors perform to create the initial training data. Step two, AI model A statistically averages that data to synthesize a fake voice. Step three, instead of playing that fake voice for a human, AI model B is brought in to listen to AI model A until the engineers of model A sounds like a real human. That is exactly the architecture of the voice MOS system. That is like a student writing a test, grading their own test, and using an answer key that they hallucinated themselves.
Pretty much. The human being, the only entity that actually possesses a voice and knows what one sounds like, has been entirely removed from the loop of evaluating what sounds human. It is a closed mechanical echo chamber. The AI ear is validating the AI voice. King uses this to expose the absolute absurdity of techno -utopianism. The machines are just talking to each other, optimizing for mathematical proximity to an old data set while completely detached from biological reality. It's a snake eating its own tail. Why would any rational engineer do this? It makes absolutely no logical scientific sense if your end goal is to create authentic human sounding speech that resonates with actual people. You're right. It makes no scientific sense at all, but it makes perfect unassailable economic sense.
And this is where Kang ties the entire book together. He brings all of this back to the economics of the tech industry and specifically the mechanics of venture capital funding. Okay. In the absence of real grounded scientific benchmarks. Economic viability becomes the benchmark. Now the puzzle pieces are fitting together. Kang outlines what he calls the three P's that drive this entire ecosystem, progress, profit, and promise. The VC hype machine. Exactly. Think about companies like 11 Labs. Investors are pouring hundreds of millions, even billions of dollars into these voice AI startups. These massive valuations are not based on the scientific truth of the product.
They aren't based on a deep fundamental understanding of human acoustics or biology. What are they based on? They are driven entirely by the promise of scale. The promise that you can mass produce the human voice as a cheap, infinitely reproducible commodity. And to keep that VC money flowing, you have to constantly show charts going up and to the right. You have to show progress. Yes. And when an AI can automatically evaluate another AI, you can run millions of tests in a single day. You can iterate and scale infinitely faster than if you had to wait for pesky slow human listeners to give you a mean opinion score. Wow.
In the tech world, algorithmic speed and scale equal perceived progress. Perceived progress generates hype, hype generates VC funding, which equals profit. The financial imperative to scale quickly completely overrides the scientific reality of the technology. The machine evaluating the machine is just a mechanism to keep the money moving. So, after unpacking all of this, what does it all actually mean for us? We've traced this incredible, somewhat terrifying arc through Kang's book. We started in the 18th century with Jacques de Volcansant's mechanical duck, creating his surface -level illusion of biology to fool an audience. Right. Then we moved to the 1960s with the FBI and Lawrence Kirsta trying to turn messy voices into neat visual fingerprints and failing miserably because they ignored the environment and the equipment. And from that failure, we saw the pivot to deep learning, the total abandonment of causal science for statistical correlation, leading to systems like Batvox, where the human subjectivity is just buried under complex math.
Right. And then we saw how that mathematical flattening is applied to everyday life. We have AI systems that are actively judging our personalities in our team fit and job interviews, purely based on our acoustic proximity to the people a company already hired. We have emotion AI policing our happiness in call centers, forcing people to sound like exaggerated cartoons or risk losing their pay. We saw the severing of the voice from the body entirely. cloning actors without permission and stripping accents in real time to create a mathematically generated shield of whiteness.
And it culminates in the tech industry setting up systems to evaluate themselves, completely removing the human ear from the loop, all in the name of venture capital scale. Tying all of these disparate threads together is the overarching thesis of Kang's work. The book serves as a massive warning about the danger of using the computer as a metaphor for the human mind and the machine as a metaphor for the human body. Knowing the voice through artificial intelligence does not mean we have achieved a deeper, more profound understanding of humanity. Not at all. It means we have successfully rendered the voice mechanical.
We have stripped away the physical body, the cultural context, the genuine emotion and all the beautiful messiness of human interaction, just so we can force the voice into a high dimensional vector space, turning it into something measurable, predictable, and ultimately, sellable. We're building these incredibly powerful, globally deployed systems based on a highly flawed, deeply flattened metaphor of what it actually means to be a human being. And that brings us to a final pretty chilling thought to leave you with as we wrap up this deep dive. Yeah. If artificial intelligence inherently relies on flattening our messy, beautiful, physical, embodied voices into neat mathematical averages in order to function, then the real danger isn't that machines will eventually figure out how to sound perfectly human. No. The real danger is that to survive in a world gated by these AI algorithms, whether you're applying for a job and your voice is being analyzed for fit, or you're calling customer service and trying to get the bot to understand your frustration, or you're just trying to be recognized as competent in your everyday life, we will have to start sounding more like machines.
To survive the system, we will have to adapt our own messy footprints to fit perfectly into their broken x -ray machines. Exactly. The diagnostic muddy waters aren't clearing up anytime soon. We're just learning how to drown in them quietly, altering our own voices to appease the algorithm. Keep that in mind the next time you ask your phone for the weather, or leave a voicemail you know is being transcribed by a server somewhere. Whose voice are you really using? Yours? Or the one the machine wants to hear?