Live data from Hacker News

NaturalSpeech 2: Zero-shot speech and singing synthesizers

speechresearch.github.io

11–20 of 126 posts

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#11
post #2

I tell friends that the scene below from T2 doesn't feel futuristic anymore. In fact, it now feels... almost mundane. I mean, a smart "script kiddie" with a bit of ML expertise can pull off this kind of deepfake voice spoofing on a relatively cheap desktop computer nowadays. We live in interesting times. SCENE: T-800, speaking to John Connor in normal voice: "What's the dog's name?" John Connor: "Max." T-800, imperso…

We aren't even far off from an LLM being able to infer that the parents are fake on the basis of the dog name. I'm not even gonna touch the chain gun shooting up a parking lot aspect.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#12
post #8
post #6

Transformers and Diffusion Models seem to be leading the pack lately in many tasks. It’s cool how these models can be used in a variety of quite different contexts without changing much about the network architecture. That being said, I think it is only a matter of time before cyber criminals develop an end to end fully automated penetration system that registers domain names, writes emails, makes phone calls, finds…

mostly agree I think the web is over as we know it maybe the solution will be the broken web plus some new system that has ties into local regulation ID systems so that you are accountable for your actions

That's (seriously) where blockchains can help.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#13
post #2

I tell friends that the scene below from T2 doesn't feel futuristic anymore. In fact, it now feels... almost mundane. I mean, a smart "script kiddie" with a bit of ML expertise can pull off this kind of deepfake voice spoofing on a relatively cheap desktop computer nowadays. We live in interesting times. SCENE: T-800, speaking to John Connor in normal voice: "What's the dog's name?" John Connor: "Max." T-800, imperso…

We aren't even far off from an LLM being able to infer that the parents are fake on the basis of the dog name. I'm not even gonna touch the chain gun shooting up a parking lot aspect.

We're safe for a few months. Asked about how did T-800 figured the foster parents are dead based on the exchange, and got:

"Yes, there is a strong cue in the exchange that suggests to the T-800 that John's foster parents are dead. The cue is that when the T-1000, impersonating Janelle, answers the phone and John asks about Wolfie, she responds by saying, "Wolfie's fine, honey. Wolfie's just fine." The use of the word "just" in this context can indicate that Janelle is trying too hard to convince John that everything is okay with the dog, suggesting that she may be lying or trying to cover up something. This small detail is enough to alert the T-800 that something is amiss and to determine that John's foster parents are likely dead".

Here's the rest of the conversation for fun:

Q: That's not it.

A: I apologize for the confusion. I'm afraid I'm not able to guess the cue you are referring to without more context or information. Can you please tell me what the cue is?

Q: The T-800 knows that the dog is named Max, but deceptively asks about "Wolfie", to which the T-1000 answers that Wolfie is fine. The real Janelle would have known that the dog is not named Wolfie.

A: Thank you for pointing that out. You are correct that the T-800's knowledge of the dog's name is a significant cue that suggests to him that something is wrong. By asking about "Wolfie" instead of Max, the T-800 is intentionally using a false name for the dog to test whether the person on the other end of the line is actually John's foster mother or the T-1000 in disguise. When the T-1000 responds with "Wolfie's fine," it confirms to the T-800 that Janelle is an imposter, as the real Janelle would have known the dog's real name. The T-800's subsequent revelation to John that his foster parents are dead is based on this realization and its mission to protect John from the T-1000.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#14
post #9

Some poking around the authors of the paper brought me to: AUDIT: Audio Editing by Following Instructions with Latent Diffusion Models [1] with demos [2] This sparks my interest so much since the last few days I was wondering if it was possible to use diffusion models on spectrograms to do audio effects editing. Here is a paper submitted a couple of weeks ago doing just that. And the demo examples are exceptional. I…

You might be interested in Riffusion, which was able to do text-to-audio by finetuning Stable Diffusion on spectrograms late last year.

https://www.riffusion.com/about

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#15
post #9

Some poking around the authors of the paper brought me to: AUDIT: Audio Editing by Following Instructions with Latent Diffusion Models [1] with demos [2] This sparks my interest so much since the last few days I was wondering if it was possible to use diffusion models on spectrograms to do audio effects editing. Here is a paper submitted a couple of weeks ago doing just that. And the demo examples are exceptional. I…

Doesn’t riffusion create music based on spectrographs?

https://github.com/riffusion/riffusion

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#16
post #13

Earlier quoted context omitted.

We aren't even far off from an LLM being able to infer that the parents are fake on the basis of the dog name. I'm not even gonna touch the chain gun shooting up a parking lot aspect.

We're safe for a few months. Asked about how did T-800 figured the foster parents are dead based on the exchange, and got: "Yes, there is a strong cue in the exchange that suggests to the T-800 that John's foster parents are dead. The cue is that when the T-1000, impersonating Janelle, answers the phone and John asks about Wolfie, she responds by saying, "Wolfie's fine, honey. Wolfie's just fine." The use of the word…

You have to use gpt-4 for this stuff man.

Direct Response from gpt-4:

The T-800 figured out that John's foster parents were dead based on the exchange because when it asked about "Wolfie" (a made-up name for the dog), the T-1000, impersonating Janelle, did not correct the name and instead went along with it, saying "Wolfie's fine." If the real Janelle had been on the phone, she would have corrected the T-800 by stating that the dog's name is actually Max. This indicated to the T-800 that the person on the other end of the line was not John's real foster parent, and thus, they were likely dead.

EDIT- Tried again after changing anything that would point to the Terminator franchise. Still nailed it.

Ron figured out the foster parents are dead based on the exchange because when he impersonated Harvey and mentioned the dog's name as "Rovy" instead of "Bingey," R-658 (impersonating Janine) did not correct him or question the name. This indicated that R-658 didn't actually know the dog's real name and was likely trying to deceive them. This deception, along with the concern about their whereabouts, led Ron to deduce that the foster parents were probably dead.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#17
post #2

I tell friends that the scene below from T2 doesn't feel futuristic anymore. In fact, it now feels... almost mundane. I mean, a smart "script kiddie" with a bit of ML expertise can pull off this kind of deepfake voice spoofing on a relatively cheap desktop computer nowadays. We live in interesting times. SCENE: T-800, speaking to John Connor in normal voice: "What's the dog's name?" John Connor: "Max." T-800, imperso…

We aren't even far off from an LLM being able to infer that the parents are fake on the basis of the dog name. I'm not even gonna touch the chain gun shooting up a parking lot aspect.

Not Far off ? GPT-4 already gets it.

Response after pasting in the scene: The T-800 figured out that John's foster parents were dead based on the exchange because when it asked about "Wolfie" (a made-up name for the dog), the T-1000, impersonating Janelle, did not correct the name and instead went along with it, saying "Wolfie's fine." If the real Janelle had been on the phone, she would have corrected the T-800 by stating that the dog's name is actually Max. This indicated to the T-800 that the person on the other end of the line was not John's real foster parent, and thus, they were likely dead.

EDIT- Tried again after changing anything that would point to the Terminator franchise. Still nailed it.

Ron figured out the foster parents are dead based on the exchange because when he impersonated Harvey and mentioned the dog's name as "Rovy" instead of "Bingey," R-658 (impersonating Janine) did not correct him or question the name. This indicated that R-658 didn't actually know the dog's real name and was likely trying to deceive them. This deception, along with the concern about their whereabouts, led Ron to deduce that the foster parents were probably dead.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#18
post #10
post #6

Transformers and Diffusion Models seem to be leading the pack lately in many tasks. It’s cool how these models can be used in a variety of quite different contexts without changing much about the network architecture. That being said, I think it is only a matter of time before cyber criminals develop an end to end fully automated penetration system that registers domain names, writes emails, makes phone calls, finds…

We've had the solution in the form of basic TLS cryptography and verification for decades now though, the problem is no one's implementing it. Governments already maintain registers of legally operating businesses: there's no reason that registration should not also be issuing cryptographic certificates which verify all forms of outbound communication by that business including phone calls. But despite telecom being…

The problem here is authenticating that you are talking to a human and not an AI. Certificates don't help with that.

Criminals will register temporary businesses and obtain certificates for them with no problem.

Re: NaturalSpeech 2: Zero-shot speech and singing synthesizers

#20
post #10
post #6

Transformers and Diffusion Models seem to be leading the pack lately in many tasks. It’s cool how these models can be used in a variety of quite different contexts without changing much about the network architecture. That being said, I think it is only a matter of time before cyber criminals develop an end to end fully automated penetration system that registers domain names, writes emails, makes phone calls, finds…

We've had the solution in the form of basic TLS cryptography and verification for decades now though, the problem is no one's implementing it. Governments already maintain registers of legally operating businesses: there's no reason that registration should not also be issuing cryptographic certificates which verify all forms of outbound communication by that business including phone calls. But despite telecom being…

And then the government will get hacked and the data leaked.
Post reply on HN