We recognize that generating speech that resembles people's voices has serious risks, which are especially top of mind in an election year. We are engaging with U.S. and international partners from across government, media, entertainment, education, civil society and beyond to ensure we are incorporating their feedback as we build.
People had been theorizing the slowdown in releases was because of this, I didn't expect they would come out and admit it.
I've found the quality of Elevenlabs' voice generation is at least twice as good as this, which is surprising since OpenAI has had more time and money to work on this. There seemed to be an issue with inconsistent accents in the multilingual samples.
I thought everyone (11Labs, open source, OpenAI even) already has human-level TTS models? Is there still an open challenge somewhere (e.g., is there a use case where a better model would make any difference?).
Idk training a model like this is kinda trivial for most shops. Afaik there are open source ones that do the same thing. We trained one with similar audio quality to this by accident while working on music generation.
It’s getting clearer — in the future, if you didn’t see it happen in real life then it didn’t happen.
I'm convinced that all these AI technologies have legitimate use cases. But it's far too easy to use them to generate endless streams of noise, spam, and mistrust and drown out anything valuable.
Expressivity and naturalness are still not good enough. It's getting there, but today you'd never have a voice conversation with an AI and come away thinking "boy, what an engaging, charismatic personality that AI had!" This is partly due to latency, partly due to imperfect speech-to-text which is too slow and doesn't understand conversational dynamics or nonverbal signals, but yes partly due to text-to-speech.
When it comes to naturalness for dialog, OpenAI actually has the best voices I've heard—significantly better than Azure or 11 Labs which are second and third best. But at least the voices they expose through ChatGPT are fairly muted and inexpressive.
Voice Engine preserves the native accent of the original speaker
I sort of understand this as a goal, but the American-accented German is weird. Like they nail some difficult "r" sounds but botch the easy ones. An odd hybrid between a fluent speaker and a total beginner.
Companies doing speech synthesis where one can clone a voice using a small recording should add digital watermarks into audio itself, akin to steganography. Companies that allow speech communication can check for these marks to make sure the voice was not generated. Many problems, including quality drop resistance, but this can be done and be a valid defence a year or maybe a few years. Not sure if carriers or phone manufacturers can check for these marks in default phone calls.
I think most of them do… but they are too easy to remove.
It isn’t just quality drop resistance, it is also filtering and post processing resistance. It is kinda like antivirus signatures: those who want to abuse it just filter and tweak it till it no longer triggers the watermark detection.
Yes! Your business accepts speech input. So you pay me - your average shady security vendor - a good wad of cash to use my universal watermark detector. It's guaranteed to detect AI stegotech, or your money back!
We achieve a whopping 30 coverage over known voice clone services. It's better than nothing!
Specifically, we encourage steps like:
* Phasing out voice based authentication as a security measure for accessing bank accounts and other sensitive information
I've been trying to tell people this for years. My bank still tries to get me to sign up to voice id whenever I call in. Voice is a terrible method for authentication.
People had been theorizing the slowdown in releases was because of this, I didn't expect they would come out and admit it.
I've found the quality of Elevenlabs' voice generation is at least twice as good as this, which is surprising since OpenAI has had more time and money to work on this. There seemed to be an issue with inconsistent accents in the multilingual samples.
They might have been forced to give a signal after this rose on HN today:
https://research.myshell.ai/open-voice
https://news.ycombinator.com/item?id=39861578
The quality of these samples would be more than sufficient to fool anyone I know.
I thought everyone (11Labs, open source, OpenAI even) already has human-level TTS models? Is there still an open challenge somewhere (e.g., is there a use case where a better model would make any difference?).
I haven’t seen any TTS take 15 second audio samples and capture the speech so well.
It’s getting clearer — in the future, if you didn’t see it happen in real life then it didn’t happen.
Even that...Everybody better get their own Certificate Authority and a family password.
Idk training a model like this is kinda trivial for most shops. Afaik there are open source ones that do the same thing. We trained one with similar audio quality to this by accident while working on music generation.
I'm convinced that all these AI technologies have legitimate use cases. But it's far too easy to use them to generate endless streams of noise, spam, and mistrust and drown out anything valuable.
Expressivity and naturalness are still not good enough. It's getting there, but today you'd never have a voice conversation with an AI and come away thinking "boy, what an engaging, charismatic personality that AI had!" This is partly due to latency, partly due to imperfect speech-to-text which is too slow and doesn't understand conversational dynamics or nonverbal signals, but yes partly due to text-to-speech.
When it comes to naturalness for dialog, OpenAI actually has the best voices I've heard—significantly better than Azure or 11 Labs which are second and third best. But at least the voices they expose through ChatGPT are fairly muted and inexpressive.
I sort of understand this as a goal, but the American-accented German is weird. Like they nail some difficult "r" sounds but botch the easy ones. An odd hybrid between a fluent speaker and a total beginner.
The Japanese has the same problem. It’s sounds like someone who is fluent, but maintained a beginner level accent, with a slight tinge of robot.
Companies doing speech synthesis where one can clone a voice using a small recording should add digital watermarks into audio itself, akin to steganography. Companies that allow speech communication can check for these marks to make sure the voice was not generated. Many problems, including quality drop resistance, but this can be done and be a valid defence a year or maybe a few years. Not sure if carriers or phone manufacturers can check for these marks in default phone calls.
I think most of them do… but they are too easy to remove.
It isn’t just quality drop resistance, it is also filtering and post processing resistance. It is kinda like antivirus signatures: those who want to abuse it just filter and tweak it till it no longer triggers the watermark detection.
Yes! Your business accepts speech input. So you pay me - your average shady security vendor - a good wad of cash to use my universal watermark detector. It's guaranteed to detect AI stegotech, or your money back!
We achieve a whopping 30 coverage over known voice clone services. It's better than nothing!
I've been trying to tell people this for years. My bank still tries to get me to sign up to voice id whenever I call in. Voice is a terrible method for authentication.