Radio host David Greene says Google's NotebookLM tool stole his voice
washingtonpost.com143 points by mikhael 13 hours ago
143 points by mikhael 13 hours ago
https://archive.ph/p81wc
Compare for yourself. David Greene: https://youtu.be/xYxQrLp4MQk NotebookLM: https://youtu.be/AR4dRtzFvxM I think he just has "podcast guy" voice. It's pretty generic. Yup, it's absolutely not his voice. The NotebookLM voice is pitched significantly higher. Nor does it seem like his voice but changed "just enough" (like in pitch). I agree, he just has a very generic-sounding "podcast guy" voice. And obviously, NotebookLM trained on tons of podcasts and is generating a highly generic, average-sounding voice. Which is why it's pitched higher, since David Greene has a lower than average pitch. This lawsuit is either just to generate buzz to build his personal brand, or maybe he's worried about the competitive threat from AI. But there's no way he's going to win this suit. This isn't like the case with Bette Midler, where Ford intentionally hired someone to mimic her voice. >Yup, it's absolutely not his voice. However it does seem to copy the the way he "lisps" his S's. I am not sure that is common 'generic-sounding "podcast guy" voice'. He almost certainly doesn't "lisp" his S's if you heard him in real life, but says them perfectly normally. What you're hearing is the way microphones deal with the hissing of an "s", same as they struggle with plosives like "p", from the whoosh of air. It's an artifact of microphones close to the mouth, so it makes sense that Google replicates it. You can use physical pop filters or digital audio filters to reduce the effect, but podcasters don't usually use the physical ones, and the level of audio processing podcasters do really depends on their level of expertise and how much they even care. I'm glad this comment was here, it was the first thing I latched on to that seemed very specific to this person (or at least uncommon amongst general "podcast guys"). In particular, check out the pronunciation of the trailing S is the word "this" at 28 seconds in the clip of Davide Greene compared to 24 seconds in the Notebook LM clip. Really seemed uncannily similar to me. When I was university, I got heavy into spoken word poetry and making hip hop. Personal Journals by Sage Francis had become a huge part of my life and rather than focus on school like a mature person I thought I’d make my own album. It turns out I have a profound lack of talent, but I got a lot of experience with a microphone. To me that ‘s’ sound reminds me of the sibilance a Shure SM58 picks up without a pop filter. I hear a different side of the same idea on ‘p’ and ‘b’ as well. I had a speech impediment as a youngster and the sound got in my head. Now I hear it on podcasts. You never know, it might be worth a couple hundred grand in settlement money… It doesn't matter if the voice is a perfect facsimile — it only matters whether a court can be persuaded that the result is derivative. As the article notes, the AI doesn't even have to be trained on Greene's voice for him to have a case. > Grimmelmann said Greene doesn’t necessarily have to show definitively that Google trained NotebookLM on his voice to have a case, or even that the voice is 100 percent identical to his. He cited a 1988 case in which the singer and actress Bette Midler successfully sued Ford Motor Company over a commercial that used a voice actor to mimic her distinctive mezzo-soprano. But Greene would then have to show that enough listeners assume it’s Greene’s voice for it to affect either his reputation or his own opportunities to capitalize on it. In other words, can you guess who someone is impersonating even if their impersonation isn't a perfect simulation? There's a lot of characteristics to people's voices. Tons of people impersonate Trump purely through cadence. Same with Obama. How many singers impersonate Tom Waits? I hear this one. I tend to catch patterns in tempo as much or more so as timbre and this is awfully close on both accounts. I don’t hear the Chris Fisher comparison that was also posted. So I would say that where there is smoke there is sometimes fire at this point. It sounds similar, but doesn't sound the same to me.
Also how would you determine the similarity allowed? Maybe if we would have such a measure they could use that in voice model training to not allow that much similarity to a single voice, but if we don't have an agreed upon value for that than it's a subjective "sounds the same to me" rule then it's hard to follow that.
Ok, they can say that don't train on their voice, but it's very likely that a blend of voices from an "allowed" set could produce a very similar voice to his. Yeah, this shouldn’t even be on HN, or Washington Post for that matter. There are going to be countless people that think AI is using their voice. Humans share remarkably similar voices, but obviously you can’t copy that (other than impersonations, obviously). Unless there is evidence that a company intentionally went after a specific human voice to train their AI, there’s no reason to report on these people claiming AI is using their voice. Maybe if it’s someone with a very distinctive voice. But this guy, as the OP said, just has a “generic podcast guy” voice. You absolutely can copy that, it’s called voice cloning and you can do it on as little as a few seconds of audio. Once cloned, you can generate audio with that voice, saying whatever you want it to. To be clear, I mean someone can’t file a lawsuit against someone else for sounding like them. Of course you can have an AI target someone else’s voice. My point is that unless there is evidence it was intentional, it’s silly to claim that just because it sounds similar to a human’s voice, that means it must’ve been intentional. I mean someone can’t file a lawsuit against someone else for sounding like them. But they did. It's literally what the article, and this thread are about. This is about an AI sounding like a human, not a human sounding like a human. The more familiar you are with his voice the less similar it would sound. It’s like how siblings look more similar to strangers. Well remember that how your voice sounds to you isn’t what other people hear. But I’m the guy who blurts out how the voice actor for the gate guard played the brother in that movie with that guy. And I can hear what he’s complaining about. There’s a lot of elements of his voice and the tempo is pretty close. )usually it’s the tempo and certain phonemes that give people away to me when they are doing a different accent) I've listened to tens of hours of NotebookLM, and this doesn't even seem close. If someone had played his voice for me and asked if it sounded like any LLM/bot I was aware of, I would have said no. It would not have even occurred to me that they were thinking of NotebookLM. As @crazygringo said, David's voice is lower. I think it might have some of the same harmonics, but it has some lower ones too, which make the overall sound come across as lower-pitched. I'm not using technical terminology here, so perhaps someone can jump in with the appropriate terms. Probably it's what his voice sounds like to himself. Maybe he should listen to his own podcasts more. I thought about this. IIRC a person's voice sounds deeper in his own head because it reverberates in his skull. if anything, this would make David think the NotebookLM voice sound less like his (since his is already deeper, to begin with). But it's possible that a professional radio host is aware of this effect and adjusts for it when deciding if something sounds like him. Maybe he over-adjusts and thinks this is what he sounds like. As for his wife, it's possible that he speaks in a higher/friendlier register when talking to her/their kids. When I tried NotebookLM on a long project management training deck, I thought the male voice sounded quite a bit like Leo Laporte. The format and banter seemed similar, too. Wow, I haven't heard that name in a long while. Brought me back to watching he and Kevin Rose on TV after school. Congratulations. I hate both of them. Maybe I’m old but the podcast style of “there might be some interesting information here, but let me tease it for ages with a voice that makes you think something interesting is about to happen…” No sir, I don’t like it This is the same reason I’ve watched about 4 hours of YouTube since it launched. Almost all car-repair videos. I watch a lot of synthesizer videos, and over the years an wholly organic 'no talking' genre has emerged for just this reason. Some people do reviews via subtitles. Probably an unpopular opinion on this forum where everyone is considering can something be done vs should something be done, but it sounds like theft to me. But I am also very anti-AI in the artistic space, because if it weren’t for humans freely providing so much artistic content, we wouldn’t have this outcome. And I believe the only end result will be less humans openly sharing knowledge, because some heavily money backed entities will just steal all the art and put it behind a paywall or advertisement. As much as I appreciate the easy search (because actual useful search has become nonexistent since AI) and the ability to ask AI to find some metadata from a large data payload, I also dislike AI, because it has effectively broken the open internet and the willingness for humans to be open to freely sharing knowledge. It's not theft, it's copying. Two different words, with two different meanings, and different legality, for very good reason. You can only steal things that can be taken away, which is why theft is bad, because it deprives the original owner of something they once had. Copying does not directly deprive anyone of anything. In fact it just adds more value to the world, and makes it more available to more people. Nobody can "copy" stuff and put it behind a paywall, because the original is still free. It's the prevention of copying that leads to expression being locked behind paywalls. It's said that copying disincentivizes creativity and creation, but in practice it does the opposite. Just look at the incredible amount of music, fiction, software, stories, art, and information that have proliferated since the birth of the web. What copying does do is it indirectly deprives people and companies of the ability monopolize profits on particular expressions without competition. But I'm not so sure that's a bad thing. For example, look at the software industry. I'm extremely grateful that patents and copyright are so rarely enforced in software and UI design, and that we've all been copying the good ideas that came before us for decades with no consequence. I'm grateful the same is true of food recipes, too. I think the world would likely be a richer one if this was true for most fields and art. That is the best explanation I have seen yet of the difference. I'm definitely stealing it... Back in the day there were these Star Wars games. Now obviously Mark Hamill costs money and he wasn't going to come back for anything less than a Disney "offer you can't refuse" pay check. So they got someone who could fake it pretty well. Ofcourse fast forward in 2026 an actor automatically sells off their face, voice and soul when they sign a contract in perpuity. There seems to be a lot of confusion in this thread around the human mind's processing of voice sounds. As with most (all?) things we do, exposure is king. This is how we don't die from trying to process infinite dimensional reality. The brain compresses, it prunes. Things seem similar if you don't have much need to distinguish them. Unless you've listened to hours of either NotebookLM or Greene, you simply won't be able to participate in the distinguishing of these voices with much ability. I listen to some Jupiter Broadcasting podcasts. The main host (Chris Fisher) regularly pops up in NotebookLLM content, with his voice. Sometimes it just jumps in, and then after some time out again. It’s usually a pretty perfect imitation, I can’t hear the difference . Edit, here an older piece, there have been many since: [0], it’s the 3rd voice that enters the NotebookLLM clip so it takes a minute before it comes in (shared this clip here late 2024 [1]). [0] https://podverse.fm/clip/Vy4y7ZG2Rd [1] https://hn.algolia.com/?query=NotebookLM%20Copied%20a%20Podc... Yeah, I don't hear it. I kept listening waiting to hear the voice that was supposed to sound like him, and never did. Was it the first one (I heard three different voices during the clip)? That one is considerably deeper than the podcaster's voice, and has different tones, too. It definitely wasn't the last one, that one was much higher pitched (and then a female voice in the middle). Feels like a big stretch, to say the least. But I can tell a big difference between the two. Ultimately, it's like some of the music copyright lawsuits, where they're suing over chord progression. There are a billion voices on the planet -- any AI generated voice is going to sound similar to someone else's real voice (and again, I don't hear it at all in this case). EDIT: So it's the third voice apparently. The pitch is close, but the tones and accents still definitely feel "off" enough that it doesn't sound like they were intentionally going for this guy. It still feels like a stretch to me, but not as much as the first voice did. It’s the voice after the woman indeed. I think it’s very close, didn’t understand what happened the first time I heard it. And this was 2024, they found many funny examples and they get better and are even better copies. And presumably anyone who has had vocal coaching for speech is going to sound somewhat similar to whatever is considered "normal" for where they live. I think I rememeber an episode where he played a clip of AI Chris talking about Linux at the start of an episode and I genuinely couldn't tell the difference In the clip, I thought he was playing a prank by reading the script of NotebookLM as the third voice (after the woman). Was that really NotebookLM? I've only heard the first two voices and the first voice didn't sound like him to me, but the last one definitely sounded like him. Yeah it’s after the woman enters. That is usually how it happens, suddenly his voice comes in, even though it’s a duo suddenly it’s his voice for some time. And really with all his mannerisms. I guess there is just a lot of his material out there. I don’t agree with this one, which puts me at one yes and one no. But it is always possible that this is what Chris sounds like in his own head. Nobody listening to audio will hear it the way he does. I know someone who is one of the top-rated voiceover artists on Fiverr. Their voice appeared on a video recently I heard and it was definitely cloned from them (as it was not a project they would have worked on). It's so easy to do now. You can just grab your favorite voiceover artist's demo reel and clone it from there. The chances of getting caught are slim, and what is the (poorly paid) artist going to do? Most of them will lack the resources to fund a protracted court case to sue some anonymous users in Tajikistan making AI slop videos en masse. People are REALLY bad at distinguishing voices. For every AI voice, there are tons of people claiming it's theirs. Well known problem for TTS. People are bad at distinguishing strange voices in a lineup, yes. That is, anyone in this thread who hasn't heard much of either the NotebookLM or Greene's voice would be a terrible witness. However, the equation changes considerably when the voice becomes familiar. You can imagine it like going from CPU to an ASIC. The brain is rather good at telling when a voice is your friend or not, the evolutionary pressure should be clear. Therefore, the people most qualified to speak on this matter will be first and foremost Greene and his podcast fans. It's a matter of exposure. I always thought the female voice sounded eerily similar to Tracy Alloway. Also available on msn.com sans paywall: https://www.msn.com/en-us/news/us/he-spent-decades-perfectin... His voice doesn't sound that distinctive to me. He's going to have a hard time unless he can find some emails that say, use David Green's voice. I think a lot of sport announcers sound the same. There might just be classes of voices where you expect a faceless voice in some scenario to sound a certain way. In general, without any context, I doubt there are a lot of people you'd immediately recognize by their voice. When I was podcasting (and editing) there were certainly some people I would recognize but in general not so much. I don't think I have a particularly distinctive voice. But I once asked a question at a Q&A with a famous actor, the whole event went up on YouTube, but only the stage was on-camera, not the people in the Q&A line. Three years later, a friend of mine messages me "WAS THAT YOU asking a question about x to actor y at event z?" She'd randomly come across it on YouTube. Yup, it was. Blew me away. > I doubt there are a lot of people you'd immediately recognize by their voice. There is a lot of variability on this from person to person. A lot of people are terrible at recognizing voices out of context. I have always been able to recognize people's voices just about as easily as their faces. (Unfortunately, while this is a neat parlor trick, I haven't found it to be a particularly valuable skill). He'll likely file in California or Federal and ask for Jury trial. I think a Jury will be sympathetic. I doubt Google will want this to go to a jury trial - not worth the risk, further news cycles of negative PR and impact on staff morale. NPR is credible and liked. > Greene’s lawsuit, filed last month in Santa Clara County Superior Court, alleges but does not offer proof that Google trained NotebookLM on his voice. His voice doesn't sound that distinctive to me. It doesn't matter whether it sounds distinctive to you. What matters is whether it's close enough to the real person's voice to be an infringement. Just like it doesn't matter if you used a machine to duplicate a painting. It's still an infringement. You can't publish a Harry Potter novel and then throw up your hands and say, "It wasn't me. The AI decided to name the characters Hargid and Hermione and Snape." Google says it paid a voice actor. If it provides proof of that, good. But like with a lot of AI things, we're in new territory here. Seems like there's a market for a tool that can compare an AI voice to a library of known famous voices so that companies like Google can tweak their machines to not sound too much like someone who can be harmed by a sound-alike. > What matters is whether it's close enough to the real person's voice to be an infringement. Also not sufficient. There has to be some evidence they attempted to copy the voice rather than just found one that was eerily similar. This comes up from time to time without AI either. Like its not good if a firm goes out to find someone with a voice similar to a famous person / voice actor…but its fine if they just randomly find one that sounds exactly the same and they say “oooh lets go with this one” and not “oooh perfect this sounds just like Dan LaFontaine!” At a certain point with generative AI we're going to run out of voices and faces the same way we run out of domain names and trademarks. Aren't these models are trained publicly available data? this might hold for some rando you doesn't have their likeness in many places to be gobbled up by the Datamongers but these programs imitating someone who has been in the media for 20 years like David Greene is not the result of chance unless you are being excessively charitable. Even if it is complete chance, there's no way to peer inside and confirm that because these things are completely opaque black boxes Can we not sample indefinitely from the latent space of vocal and delivery characteristics? the "latent space containing all voices" may give you the ability to parametrize voices and make an infinite number of unique voices. BUT... people have a limited ability to distinguish points in that space. in perceptual psychology/psychophysics, there's the concept of the "just-noticeable difference" (JND) which is the smallest change to a stimulus you can make that is reliable detectable. normally the JND is measured on physical properties like brightness, pitch, etc but there's no reason it couldn't be applied to a more abstract latent space. two points in a particular latent space may be mathematically unique, but if they're indistinguishable to humans we shouldn't treat them as distinct voices If someone is born with David Attenborough‘s exact voice… what happens? No problem. But if they train intensively to achieve that and compete with David, maybe problem? IANAL Echoes of when Scarlett Johansson accused OpenAI of stealing her voice. That time it was impossible to tell who was in the right - there was no available recording of OpenAI's supposed Scarlett clone - they had pulled it immediately for fear of bad PR. Then came the completely nonsensical HN threads with people arguing about something they hadn't heard. Maybe don't redo that whole thing? Could we at least make sure to secure some examples of A and B, this time? -- Statement from Scarlett Johansson on the OpenAI "Sky" voice (May 20, 2024) https://news.ycombinator.com/item?id=40421225 (1021 comments) OpenAI didn’t copy Scarlett Johansson’s voice for ChatGPT, records show (May 23, 2024) https://news.ycombinator.com/item?id=40448045 (1218 comments) That's not it. This is the first version of the Sky voice model which most people seem to agree doesn't sound that much like Scarlett. See my adjacent comment for more details. Don't know what you're talking about, clips of the AI voice were publicly available at the time. They were definitely not. Here is someone in that first thread trying to figure out what is actually going on, and failing: https://news.ycombinator.com/item?id=40421757 I had to wade through 12 gigantic generic political subthreads to find this. "Do you have an example of the changed voice anywhere?" (No replies.) "Yes, I feel gaslit by the whole situation" is a great summary. Please post a clip from the time. I'm still curious to hear how similar or not they acually were. > Greene felt the male voice sounded just like him Turns out he still has his own voice, that one sounds like him. That's likely the case because they deliberately cloned his voice. I've been a pro sound engineer for 30 years, doing music, film, and producing podcasts for other people, plus I have years of prior training in elocution, which requires knowing about the anatomy and articulation dynamics of the human voice. The voices don't sound that similar to me, and he has what I think of as a generic mid-atlantic accent. I'm sure it feels uncomfortably similar to him, but I think this is a mix of confirmation bias and the fact that radio and tv stations have long selected for 'average' sounding voices because listeners and viewers will call in to complain about voices they find annoying. Performers in this field cultivate those kind of generic voices, much as real estate agents cultivate aim for a friendly-but-bland look rather than trying to stand out individually. I do feel for the guy a bit because voice generation is now so good that there's no reason to pay performers with a 'radio voice' for commercial voice-over or narration work in many cases, and I question the value of applying AI to fields of personal rather than industrial endeavor - was the cost of human vocalization such a drain on the economy that we are better off for automating those jobs away as quickly as possible? However, I don't buy his claim that his individual voice and way of speaking was stolen. It's just not very distinctive to me, in the same way that faces from thispersondoesnotexist.com will inevitably approximate the appearance of some real people. If you read the article, Google says they hired a professional voice actor to create the NotebookLM voice. I'm sure this will come to light in the lawsuit. Doesn't seem like a very good clone. I wonder if he's hoping he's in their training data for a payout, if he can force that to be disclosed. I think a few random samples trivially shows NotebookLM is higher pitched, although if you generalize to "deep male voice with vocal fry" you could lump them together with half the radio and podcast voices. Unlikely. Most likely is that they used a lot of his podcasts in training and the AI picked a voice that was well represented in its training set because that's how it works. Nobody at Google was like "we should use this guy's voice!"
tantalor - 9 hours ago
crazygringo - 8 hours ago
mhuffman - 4 hours ago
crazygringo - 3 hours ago
hn_throwaway_99 - an hour ago
hluska - 3 hours ago
sailfast - 6 hours ago
rectang - 4 hours ago
godelski - 2 hours ago
hinkley - 8 hours ago
spyder - 6 hours ago
jader201 - 5 hours ago
reactordev - an hour ago
jader201 - an hour ago
reaperducer - 41 minutes ago
jader201 - 19 minutes ago
koolba - 8 hours ago
hinkley - 8 hours ago
apparent - 7 hours ago
amelius - 7 hours ago
apparent - 5 hours ago
Zigurd - 8 hours ago
kylecazar - 8 hours ago
tbossanova - 7 hours ago
irishcoffee - 6 hours ago
anigbrowl - 5 hours ago
disposition2 - 7 hours ago
csallen - 7 hours ago
alvah - 6 hours ago
expedition32 - 7 hours ago
Unearned5161 - 27 minutes ago
teekert - 9 hours ago
jader201 - 9 hours ago
teekert - 9 hours ago
ghaff - 7 hours ago
walthamstow - 9 hours ago
allenu - 9 hours ago
teekert - 9 hours ago
hinkley - 8 hours ago
qingcharles - 22 minutes ago
neural_thing - 3 hours ago
Unearned5161 - 38 minutes ago
prodigycorp - 9 hours ago
dpe82 - 9 hours ago
johnwheeler - 9 hours ago
dehrmann - 9 hours ago
ghaff - 7 hours ago
crazygringo - 3 hours ago
georgemcbay - 6 hours ago
fhub - 9 hours ago
apparent - 7 hours ago
reaperducer - 9 hours ago
nerdsniper - 9 hours ago
j-bos - 9 hours ago
b00ty4breakfast - 8 hours ago
kelseyfrog - 9 hours ago
parpfish - 8 hours ago
bethekidyouwant - 8 hours ago
recursive - 6 hours ago
lysace - 9 hours ago
Leynos - 7 hours ago
lysace - 7 hours ago
ghostly_s - 9 hours ago
lysace - 8 hours ago
barbazoo - 9 hours ago
13415 - 9 hours ago
anigbrowl - 5 hours ago
golfer - 8 hours ago
9x39 - 6 hours ago
IshKebab - 8 hours ago