Launch HN: Play.ht (YC W23) – Generate and clone voices from 20 seconds of audio
Today, we are excited to share beta access to our latest model, Parrot, that is capable of cloning any voice with a few seconds of audio and generating expressive speech from text.
You can try it out here: https://playground.play.ht. And there are demo videos at https://www.youtube.com/watch?v=aL_hmxTLHiM and https://www.youtube.com/watch?v=fdEEoODd6Kk.
The model also captures accents well and is able to speak in all English accents. Even more interesting, it can make non-English speakers speak English while preserving their original accent. Just upload a non-English speaker clip and try it yourself.
Existing text to speech models either lack expressiveness, control or directability of the voice. For example, making a voice speak in a specific way, or emphasizing on a certain word or parts of the speech. Our goal is to solve these across all languages. Since the voices are built on LLMs they are able to express emotions based on the context of the text.
Our previous speech model, Peregrine, which we released last September, is able to laugh, scream and express other emotions: https://play.ht/blog/introducing-truly-realistic-text-to-spe.... We posted it to HN here: https://news.ycombinator.com/item?id=32945504.
With Parrot, we've taken a slightly different approach and trained it on a much larger data set. Both Parrot and Peregrine only speak English at the moment but we are working on other languages and are seeing impressive early results that we plan to share soon.
Content creators of all kinds (gaming, media production, elearning) spend a lot of time and effort recording and editing high-quality audio. We solve that and make it as simple as writing and editing text. Our users range from individual creators looking to voice their videos, podcasts, etc to teams at various companies creating dynamic audio content.
We initially built this product for ourselves to listen to books and articles online and then found the quality of TTS is very low, so we started working on this product until, eventually we trained our own models and built a business around it. There are many robotic TTS services out there, but ours allows people to generate truly human-level expressive speech and allows anyone to clone voices instantly with strong resemblance. We initially used existing TTS models and APIs but when we started talking to our customers in gaming, media production, and others, people didn't like the monotone robotic TTS style. So we doubled down in training a new model based on the new emerging architectures using transformers and self supervised learning.
On our platform, we offer two types of voice cloning: high-fidelity and zero-shot. High-fidelity voice cloning requires around 20 minutes of audio data and creates an expressive voice that is more robust and captures the accent of the target voice with all its nuances. Zero-shot clones the voice with only a few seconds of audio and captures most of the accent and tone, but isn’t as nuanced because it has less data to work with. We also offer a diverse library of over a hundred voices for various use cases.
We offer two ways to use these models on the platform: (1) our text to voice editor, that allows users to create and manage their audio files in projects, etc.; and (2) our API - Personally, this sounds like an extremely irresponsible startup, but I don't know much about it so I'm trying to reserve judgement. I'm mildly agitated by your comment, so I'd like to take the liberty of pointing out that you are the only one in this entire thread linking the names Hammad and Mahmoud to racism. Everybody else in the entire thread is talking about the product on its merits. There's a heated debate and nobody gives a fuck about where the founders are from. That's how it should be. Stop making the world a worse place than you found it. And, FWIW, I think that the product looks pretty neat. And that the voting ring was just too obvious a play :-) What is your process for verifying consent? Probably not something we'll get to hear as part of the PR pitch. Or is the consent statement the thing that will be cloned and is there no separate training audio? Then it might actually work and you'll just have to get close enough that the human you're trying to fool can't distinguish anymore (defeating the need for this tech in the first place, at least in targeted rather than automated cases). I like your idea of just training on the consent text! That wasn't the case when I tried it as you needed around 3h (optimally) of training data. If you mean the white noise, I meant that as a brute force attack because, to do it more targeted (to know what it'll accept as seeming like your target voice), you'd likely need their exact model rather than doing your own. If your homepage is toadying up to Musk by claiming he has "limitless intellect" then I've already heard enough. We have a duty to consider how what we build can be used to harm others. If the obvious and many ways this could be abused aren't covered anywhere I can find on your website, then I'm going to conclude you haven't considered them. Which is terrible. Pops a modal: Try Voice Cloning for Free! Enter a credit card for $0.00/mo with no other information on screen Bounce. Why not let me play around with it a little without asking for a credit card? "no matter what based on your usage/you are locked into the free tier" would have helped for sure i still would've bounced because i just wanted to goof off with it quickly while it captured my attention and requiring a payment method is just... terrible friction to capture users being able to quickly test one of the key features you advertise, but i guess if fraud concerns are that bad, that's the tradeoff you have to accept? It's cool tech, yes I'm impressed at the achievement. Nuclear weapons are impressive too. OTOH this kind of thing is getting easier and easier to do, so what's a realistic way forward? I truly do not understand people like these founders, obviously they understand the future they're creating. "If not us, someone else would do it" is not an excuse. Neither is "I like money". Next, normal people will adopt the Mafia's 'cover the mouth while pretending to use a tooth-pick whilst talking to prevent lip reading from remote viewers (same thing sports people do currently. - My granmother was deaf for the latter half of her life. She became an expert lip reader. It was fun going to restaurants with her as she would tell me what people at the tables far away were talking about "oh that couple isnt having a happy time..." It seems a bit weird to me though. I mean, looking back at old records can still pass as mere nostalgic behavior. Wanting new sentences pronounced in disguise of lost relative voices is not great in term of respect for these people to share my own feelings. Also I guess that now there is not much preventing completely new songs with whatever lyrics staring voices of Elvis, Hendrix and Pavarotti. Actually a continuous flow of on the fly generated lyrics seems perfectly plausible at this level, isn't it? 2. Requires internet 3. What image do you verify? Between auto-retouching, manual retouching, compression, filetype conversion, an image file might be invisibly transformed 10 times in between capture and Instagram upload. 4. Useless for disproving fake images until every camera manufacturer in the world has implemented this. 5. Hostile to customers, now your picture doesn't get the green verified badge or whatever if you decide to crop it or something. This is a general problem that I have with a lot of blockchain ideas. For example, there are a few startups that claim to verify carbon offsets by registering them on the blockchain. There are many problems with this that we don't need to get into, but the relevant one here is: what is stopping me from registering the same offset on three different chains? Pretty sure this is not a good use of blockchain, and I don't see how it would remotely avoid misuses. This seems to be the only realistic future. This sort of technology literally makes it impossible to trust anything electronic. People were worried about the balkanization of the internet, but now they look like optimists. we said: "No, because there is also technology that makes it possible to trust anything electronic with very nearly 100% reliability. But no one uses it" I think your first statement is both technically wrong and generally wrong. Electronic trust is a solved problem...it's just that right now, it's really not as big a deal as some people are worried about it being, so we haven't generally implemented the solution. We could make electronic cars for long time before other things made them commercially viable. I disagree that electronic trust is a solved problem. It is mathematically solved, yes, but the reason that it isn't widely used is because it's still intrusive and painful to do. A solution that isn't acceptable to the masses isn't an effective solution. If it could be done in a way that is invisible (like HTTPS, for instance), then it would be ubiquitous. That's the part of the problem space that still needs resolution. I send you a bunch of bytes signed with my private key (which somehow you have to verify in a trusted way) and you can be sure that I am the person who signed those bytes (unless I was compromised). USB etc hardware keys can be used protect accounts like Gmail, Coinbase, web hosting services, many more now. Fun fact: it's possible to receive Facebook user notification emails encrypted against your public PGP key. Example: https://blog.elevenlabs.io/enter-the-new-year-with-a-bang/ Funny how computing perfected communication and ultimately will undermine itself. > I don't know what the way forward is for news. I'd say every packet of voice/img will have to be signed by the recording device and checked at rendering time. > I truly do not understand people like these founders Me neither. Don't do it. We don't need that, and the malevolent use of this will confuse people to an extreme point. Even the 'good' use of having a deceased relative utter new sentences is beyond strange. This is too far gone. And I'm no luddite. Apply your same logic to any other easily misused tech: "We must all have easy access to bio-engineered viruses. Otherwise only..." "We all need to have access to nuclear weapons. Otherwise only..." Not all tech should be in everyone's hands. That is not possible with or comparable to, things such as bioweapons. The difference between "the paper is out there" and "there's a button to do this" is quite obvious in cases like software exploits. A report of finding a vulnerability rarely leads to a massive automated exploitation campaign, but if that report also contains a proof of concept the amount of automated attacks radically increase. I believe the same is true for many other types of crime: even a mild bar to entry will prevent a significant amount of criminals from advancing their techniques. I think the negative impact of these voice changers is much bigger than the advantage we gain as a society. Criminals will always exist, even crafty ones, but "we can't prevent crime so let's not bother trying to do anything about it" is not a great take in my opinion. TOTC is about resource depletion. GYI. It's not applicable here. You could of course clone a voice to generate that "consent" -- but at that point there's no additional harm done because they'd already have the clone. It's unrealistic that this tech won't exist somewhere, even if the big actors stay away for ethical reasons. A voice auth practice strikes me as a good compromise. >This is the kind of thing that should be illegal. Now, any Plebian could essentially write a letter to anyone, impersonating anyone. Forged letters could drag us into a war with Persia - for Jupiter's sake! I suppose if we approach the point that we can create robotic clones of anyone, anywhere, that look, sound, and move like anyone on the planet, that will be just like the post office too, right? The analogy for this technology would be a robot that can perfectly imitate someone's handwriting and vocabulary using one letter as a reference. https://www.uspis.gov/history-spotlight/history-of-the-mail-... Maybe we'll need a new specialized law enforcement agency like the Postal Inspectors to deal with the inevitable wave of AI-assisted crime. Sure, this is marketed as generating your own voice to read scripts for you YouTube channel, but are they actually verifying who's voice you're generating? Try to deceive people by learning about their contacts, writing a convincing letter and sending it. How long does it take you to prepare one letter? Now those AIs potentially allow you to just generate millions of those with one click. The problem is the scale: everyone can do it at no cost and at scale. Not true. Making it illegal wouldn't make it nonexistent -- that's true. But making it illegal would provide at least some method of mitigating some of the harm. That's more than what we have right now. > We have to accept that this tech exists and there will be both positive and negative outcomes from it. Of course. But that doesn't mean it's futile to try to reduce the negative outcomes. It's already illegal to impersonate someone to steal money or scam them, and those laws were on the books before computers existed. > Of course. But that doesn't mean it's futile to try to reduce the negative outcomes. You can run something on a consumer GPU and it's every bit as good if you know how to dial it in. By the end of the year you'll be able to download a nicely packaged "voice cloner" from a torrent that runs on a cheap laptop. IMHO any effort on regulation is far better spent informing people rather than trying to put the cat back in the bag. I don't think so at all. There are all sorts of things you can technically do with ease that are illegal for good reason. Laws against them aren't futile. But I admit that perhaps I'm being overly optimistic here. I'm just trying very hard to see any way that this stuff can end up not being a complete societal disaster. There are two hurdles a criminal has to get past: 1. decide to break the law
2. figure out how to pull off their scam It sounds like you're saying that since hurdle #1 already exists, hurdle #2 is irrelevant? No, of course it isn't. That's like saying that gun control can't possible help because it's already illegal to shoot someone. Adding difficulty to a crime reduces (but does not eliminate) the prevalence of the crime. > The FTC Act’s prohibition on deceptive or unfair conduct can apply if you make, sell, or use a tool that is effectively designed to deceive – even if that’s not its intended or sole purpose. It seems like an awfully broad rule? But they probably could go after this startup if they noticed it. There are some kinds of businesses where making sure the regulators like what you’re doing is pretty much a prerequisite. On the other hand, plenty of companies got where they are today by pushing the limits. [1] https://www.ftc.gov/business-guidance/blog/2023/03/chatbots-... Wow. I would have imagined an article from the FTC to be more... Bland, for want of a better term. Clear, direct, confident, not overloaded with qualifiers, not afraid of metaphor, self-summarizing, signposting, and most importantly it always has an energy of some kind that government communication (in seeking to appear neutral) regularly lacks - having that energy is why it doesn’t feel “bland”. I wonder if they have internal documents to guide their writers, or if it’s mostly information stored in the heads of Lesley Fair and Michael Atleson (who between them seem write most - all? - of the posts). If other people continue to have access as well as the '3 letter agencies', the same power will still exists for the agencies, except that there will also be an essentially unlimited number of other people who could be used as scapegoats. If only '3 letter agencies' have access, they would obviously be the first ones to come under scrutiny if a case of misuse were discovered. Without that, we'll just never know, and Joe Blow who never saw a deepfake of Joe Biden praising the stickiness of the latest OG Kush will trust anything. Just like with image generation models, this will massively raise the bar for what amateurs can do with a limited budget and limited time. It's hard to justify spending thousands of dollars on voice acting and art for a hobby project, but now amateurs can get something that is 90% there and substitute professional work if the project takes off. The reason is that if you speak with lots of verbal fillers, that's actually an important part of how you sound to other people. It makes sense to clean up audio for a podcast, but not for your great grandchildren. A voice cloner doesn't care that you say "um" too much. It's parsing audio for phonemes. You can use Descript, CleanVoice, and other tools to achieve exactly what you just said, in a few minutes, from just the original recording. But yes, it would be weird to generate more stuff spoken by your father by using this technology. And beyond that, what's even the point? It's not your dad. Yes in a few years you would be able to generate a complete avatar of someone, but it isn't them, and i think it will mess with you mentally. I happen to be someone who believes that it's wonderful your dad left you with this artifact. It was a touching sentiment then, and now it can serve his obvious purpose many times over. He didn't record himself as a side-effect of disease, or because he loved that particular story in the sound of his voice. He wanted people in the future to be able to hear what he sounded like! Given that he could not have foreseen voice cloning (and therefore not explicitly asked for it) I cannot think of a more obvious example of someone wanting their voice to survive them. I wish more folks would record The Night Before Christmas. This makes for a rather explosive combination. I could, as some tobacco exec, hire some mimic to promote cigarette sales using a celebrity's distinct voice. By the time the regulators catch up, the spot has aired a few million times & made a potload of money. A bunch of celebs[1][2] have trademarked their voice...but enforcement is spotty. [1] https://economictimes.indiatimes.com/news/new-updates/amitab...
[2] https://www.financialexpress.com/archive/when-celebrities-se... whilst this is cheap and easy - increasing the potential for scams in big way - even to the point of automating the scam 1) The impersonation can be carried out in real-time by the criminal themselves. No need to employ anyone else. (No trail leading to them.) 2) Pro impersonators aren't common in society. They are limited as an asset and not duplicatable. So, using one cannot spread like wildfire and overwhelm our awareness that voice impersonation is something of a common risk. Maybe the second could hold the first in check. I think disruptive tech like this & similar advances in visuals come with a societal impact that lessens potentials for realising the bigger fears.
But people just love fears. Compare: 1) Use lots of time to find person who can impersonate specific other person. Unless you give them a lot of money or threats to shut up you can't use them to real time decieve someone. 2) Clone 1 million voices from tiktok in 1 minute. Contact 10 million relatives with a synth voice that is intelligent enough to answer questions. We will have billions of AI's, containers, programs, agents running around trying to deceive absolutely everyone and their grandmother 24/7 soon. Of course a human could do that manually, but with those AIs it's a completely different scale, and it can be automated (so someone with no skills can click a button "generate 10k fake identities online"). Maybe even with one click, those techs could generate a fake coworker and send phishing e-mails. Suddenly every single e-mail you receive (or friend request or call) could be a very convincing fake. You don't have to be a high-value target anymore, it's all automated. That makes it much more dangerous: from "can be forged manually with time and resources" to "everyone can do it at scale for free". Why should the ability to impersonate a persons voice suddenly become a crime in itself? Should we arrest Jim Carrey? Isn't it when the thing was used to do something else illegal when enforcement is required? It became more apparent to me how icky this is as the voice actor of one of the most iconic characters in the game died suddenly 10 days ago... Yes, I agree, only criminals should be allowed to freely run it. We have these measures in place and are working on others to make sure the technology is used towards the betterment of humanity. 1/ Auto moderation on text to block harmful/malicious speech.
2/ As someone pointed out in the comments, we had a manual review process in place where the user is required to read out a consent and a member from Play.ht would review it before approving the voice. We're working on improving and adding this back.
3/ The user facing service is paywalled so we don't allow everyone in.
4/ Users trying to create malicious content are flagged and reviewed.
5/ A classifier to detect AI generated speech - We built and are offering for free a tool that can identify AI generated vs human-generated audio (https://play.ht/voice-classifier-detect-ai-voices/), we will continue to invest in this tool, and we hope it helps with deploying this technology safely. - If we get any reports of a cloned voice without consent, we block the user and remove the voice instantly. - The price of high-fidelity voice cloning is too high for scammers to use at scale; we have been live with it for four months and haven't had any cases of abuse so far. Like any technology, it has the potential to be abused, and we are working hard to mitigate that and deploy it safely. We will continue to observe the use cases and user feedback and improve the safety of the service accordingly. Since we launched voice cloning 4 months ago, we have seen enough genuine use cases which motivated us to keep moving forward and figure out safe ways to make the technology useful for all. How do you know this? This won't be the problem. My voice calling my parents asking for money to be sent to a random account will be the problem. And none of that will be sexual, offensive, racist, or threatening. >we are working hard to mitigate that and deploy it safely. How? >we have seen enough genuine use cases What? This is exactly what makes me so angry about "AI safety" initiatives: they are largely worrying about the wrong thing. People have been so focused on the "this may make some obscene joke, or be biased against some skin colors" that they have completely missed out on the much more serious harms that AI will cause with respect to, in this case, impersonation scams. Congrats, people can't say the N-word with your technology, but they can say "Hi Bob, just calling to verify that we did indeed change the target account where you should wire your invoice payment." All's well that ends well, though. We simply don't have the resources to continue this "breakneck pace". I also like that the cloud provider supports SSML and I can explicitly configure the emotion, whereas Playht dynamically changed the emotion based on context of the text. I had envisioned the scammers leveling up to this scam last week or so in a comment here. Though google news a few days later showed me it's already happening... https://www.dailymail.co.uk/news/article-11897239/Houston-co... Get a panicky call from "me" in the middle of the night? If I don't include my safe word, that call isn't from me. It's not very important that the voice is similar to the supposed victim. Usually the person in the call is weeping and it's very difficult to recognize the voice. Moreover a confusing voice at 2am may be interpreted as any of your relatives or friends, but an exact voice can be interpreted only as one and it's easier to know that that person is safe. He was 90% convinced that it was true, but my mother made him call me before doing anything, which saved him about $10k. She thought it was suspicious that I would have left the country without mentioning it to her. If the person he was talking to was relatively calm and sounded like me, it might have been successful. > Bububu. Hi, I'm ... bububu > John? > Bububu. Yes, I'm John. bububu. I'm in bububu ... the jail in ... bububu > Mexico? > Bububu. Yes, In Mexico. bububu. And I need money ... bububu There are other that research the victim and have more data for a targeted call, but it's more difficult so most cases where random calls where they don't have a sample voice of the victim. Their FAQ says > Can I clone anyone's voice? > Yes, we allow you to clone another person's voice if you have their consent. As you can imagine, cloning a voice which sounds exactly like the person is a powerful thing and can be easily misused. We deeply care about ethics and privacy and have implemented verfication processes and regulations to avoid people cloning anyone's voice without their consent. But I very much doubt that they've gotten the consent of even half the celebrity voices they're using to promote their service. Are you though? You might just be computer-generated. While I'm very impressed with this technically (and as a pro-audio person I feel validated to see my predictions of a few years back coming true so dramatically), I don't see anything about risk management in here. Your tech absolutely will get used by scammers, given the overabundance of voice data on the open internet. How are you going to hedge against that? I'm trying not to be reflexively dismissive, and I know the technology is evolving so fast that your individual company can't necessarily pre-empt it, any more than an email software supplier is responsible for the existence of phishing. But I work adjacent to the security space (studying violent extremists) and I can think of a ton of ways to abuse this where economics would be absolutely zero deterrent. really? that's a nice gesture… so where can I download it? Outside of memes or maybe the occasional well-intentioned prank, I really can't think of anything either. If people can find comfort hearing their mom say words of encouragement in a tough situation, I think a lot of people would do it. Kinda hard because for some others that would mean never getting closure. Weird stuff is certainly about to happen… There is vastly more potential for that to be abused by others than used in any emotionally or socially constructive way. [1] https://www.wired.com/story/a-sons-race-to-give-his-dying-fa... Mods are more difficult to attach a moral judgement to. I don't think I'd really consider them malicious, as long as they're not sold, but there's a very thin line between a high quality mod and stealing someone's voice. I can easily imagine a future where AI-generated impersonations are deemed by courts or new legislation to be protected by personality rights. In that world, voice actors could expand their business by offering deeply discounted rates for AI-generated work. Alternatively, if/when tech like Play.ht is consistently good enough, maybe it just becomes a standard practice for all voice acting work to include a combination of human- and AI-generated content, like a programmer using Copilot or a writer using GPT. No? Then why do you assume that someone else would want to do the same in their profession? As AI-generated content is not protectable under IP law, it's a non-starter for games, film, TV, or music for anything except background filler. If you had a solo contracting business, and the technology existed to fully outsource a development project to AI based on carefully documented requirements, using it would be a cheaper alternative to subcontracting. Rather than writing every line of code by hand, you would transition to becoming an architect, project manager, code reviewer, and QA tester. Now you're one person with the resources and earning potential of an entire development shop. I have my fair share of complaints about AI coding tools, but that isn't one of them. Maybe the increase in supply would result in a lower average software engineering income, but it wouldn't have to if demand kept pace with supply. Furthermore, code is more fungible than a person's voice. If someone wants a particular celebrity's voice, that celebrity has a monopoly on it. Thus, it's not obvious that increasing the supply of one's voice acting work would decrease its value. (I suspect the opposite to be the case, until a point of diminishing returns.) Although the voice acting case has a similar concern; will we get an explosion in new and/or higher-quality media, or will we see a consolidation to a smaller number of well known voice actors taking an outsized amount amount of work? Another issue, if we look beyond impersonation specifically, is that human voices may become marginalized over time in favor of entirely synthetic voices. I imagine that this would start with synthetic voices playing minor roles alongside human/human-impersonated voices, but over time certain synthetic voices would organically become recognizable in their own rights. Again, I see plenty of concerns with AI in general, but more of a mixed bag than strictly negative, and there isn't anything inherently nefarious about this product in particular. Personally, I'm optimistic about what society looks like in the long run if humanity proves to be a responsible steward of increasingly advanced AI. By the time we're at a point where 90% of people can be effectively automated out of a job, we'll have had to have figured out some alternative way of distributing resources among the population, i.e. a meaningful UBI backed by continued growth of our species' collective wealth and productivity. I can easily imagine a not-too-distant world that is effectively post-scarcity, where it's not frowned upon to spend years (or lifetimes) on non-income-generating pursuits, and where the only jobs performed by humans are entrepreneur, executive, politician, judge, general, teacher, and other things of that must be done by humans for one reason or another. So am I happy that AI is encroaching on skilled labor? In the short term, not necessarily. But it's not necessarily bad either, it's the reality that we're in, and long-term I'm more optimistic than not. Instead we will probably see licenses for generated voices. And in case for games the game developer could make the voice model freely available for mods of his game.(The mods are already using assets from the game, why not also audio?) so your logic is that all that text should be audio and people will consume more? Because I got news for you, reading is faster than listening. I think the case for having all text be listenable is pretty clear. We're all really busy and often our hands are busy but we're not doing something that mentally stimulating. This is an ideal time to listen to an audiobook, a blog, the news, or whatever else you'd like. Or, heck, my own voice. Though it'd be surreal to hear not-me-but-me saying things I've never said. But I'm not sure. Part of why I'd prefer the original author to read a book is that they vocally emphasize certain parts of the book, and I don't think these models could do that at this point. Right, but having AI read the book in the author's voice is definitely not the author reading the work. As you mention, the reason that people like to hear the author read it is because it's the author reading it, theoretically emphasizing and acting things out according to what was intended. It's not just to hear the author's voice. So I don't see what the value-add is. I'd really like something like Img2Img for voices so I can translate a performance to an arbitrary (synthetic) voice. Going to the demo page and hearing a random snippet of Musk-worship was pretty weird. Out of all audio tracks to place at the top of your demos, you chose this? Warning to others wanting to click on the link: damn that was creepy. I’m still laughing five minutes later.476 comments
[ 120 ms ] story [ 8135 ms ] thread