Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

NaturalSpeech 2: Zero-shot speech and singing synthesizers

speechresearch.github.io

Post reply on HN
Print view View original

121–126 of 126 posts

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

#121

Post by retrac »

Timon3 wrote: post #96
cloudking wrote: post #89

Earlier quoted context omitted.

My lamen understanding is that at a high level the transformer model is performing mathematical operations on the data, based on a complex series of formulas (the "model") derived from the weights set by training. Then it's able to take in new data, perform the math, and output what it thinks comes next. Is it a big stretch of the imagination to think that maybe such "models" (mathematical formulas) exist also in our…

As far as I understand this is true. I like to think about it like this: there is some magic formula f(x)=? that perfectly maps our inputs to our outputs (e.g. image captions to images, or input texts to longer input texts), but we don't know how to find it. So we build a space with incredibly many dimensions, and we learn some mapping in this space, which is hopefully very close to the magic formula. Our brains fund…

You're right about how machine learning is learning to approximate a function - most machine learning systems are mathematically equivalent to stochastic gradient descent, a statistical method which can, theoretically, do the same thing.

The surprise was that people (me, at least!) thought the computation and amount of data required to learn a function like "translate English to French" would be completely impractical to ever realize.

I think it's open question whether humans work like that, though we probably do.

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

#122

Post by AnthonyMouse »

XorNot wrote: post #111
AnthonyMouse wrote: post #88

Earlier quoted context omitted.

"Authenticating that you are talking to a human and not an AI" is just a proxy for something else, most commonly wanting to be able to rate limit something. Because otherwise, why does it matter? A lot of people are suggesting things like, have the government do it. But that has two big problems. First is privacy. If you have to prove your identity any time you want to do anything, nobody can be anonymous anymore, wh…

> If you have to prove your identity any time you want to do anything, nobody can be anonymous anymore, which is Very Bad. No one would have to prove anything. But why would I accept calls from any entity claiming to be a legally registered business which doesn't present a government issued certificate proving that? I already verify businesses by looking up business number registries. But this should be automatic to…

Now you're talking about something else. Whether the entity calling you and claiming to be your bank is, in fact, your bank is a separate issue from whether it's an AI or a human. It could be a robocall from your bank trying to upsell you on some financial product and it would still be from your bank. It could be a live human being claiming to be your bank when they're not.

And we know how to solve that one. If you get an email from bank.com, your email server knows how to verify that it was actually from the servers of bank.com, using certificates and DNS records etc. This is technology we could also apply to phone calls, notwithstanding that we haven't, and do so without any government action. Google and Apple could implement this right now if they cared to.

But that doesn't prevent spam. Because the spammer doesn't have to claim to be bank.com in particular. They can go register somethingthatsoundslikeabank.com and send their spam from there. Then you receive their spam/calls because you want to be able to receive legitimate ones from people without whitelisting them individually and the spam domain hasn't been spamming long enough to get blacklisted yet.

Real solutions look something like this: To send mail from a.example.com to user@b.example.com, a.example.com has to generate a computationally expensive hash containing their domain name and the target email address. Or transfer a few cents worth of cryptocurrency to the target in exchange for a signature of that pairing. New mail servers then have to do a lot of initial computation but once they've computed/bought hashes of 99% of the people their users communicate with they can cache the results and they're done. Spammers have to keep redoing the expensive computation every time one of their domains gets blacklisted.

Then you can AI generate all the spam you like, but if the recipients don't want it, your domains are going to get blocked, which would be expensive.

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

#123

Post by cloudking »

retrac wrote: post #121
Timon3 wrote: post #96

Earlier quoted context omitted.

As far as I understand this is true. I like to think about it like this: there is some magic formula f(x)=? that perfectly maps our inputs to our outputs (e.g. image captions to images, or input texts to longer input texts), but we don't know how to find it. So we build a space with incredibly many dimensions, and we learn some mapping in this space, which is hopefully very close to the magic formula. Our brains fund…

You're right about how machine learning is learning to approximate a function - most machine learning systems are mathematically equivalent to stochastic gradient descent, a statistical method which can, theoretically, do the same thing. The surprise was that people (me, at least!) thought the computation and amount of data required to learn a function like "translate English to French" would be completely impractica…

This is really fascinating, assuming it is true, it could imply that everything we "learn" is essentially a training process in our brain to store a new model/function. As humans we've figured out how to transfer these models between our brains through communication. Maybe it is possible to "upload" a model to the brain like Neo learning kung-fu...

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

#124

Post by woodson »

lachlan_gray wrote: post #23

Is the approach similar to what they did for Riffusion? (On front page a while ago) https://www.riffusion.com/

No, it’s not similar (other than superficially, as both use diffusion methods in some way). It uses diffusion to generate latent vectors that are entered in a neural audio codec model to produce speech.

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

#125

Post by tudorw »

hbbio wrote: post #12
tudorw wrote: post #8

Earlier quoted context omitted.

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.

quite, mentioning AI or Blockchain seem to set off a knee jerk reaction, crime is crime, methods change but frequency afaik stays relatively stable as a product of poverty Vs quality of life, so, as these technologies are still in their early days, so many factors are affecting their performance and changing their potential, seems odd that a forum for hackers is so, well, not sure how to put it, negative?

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

#126

Post by DonHopkins »

famouswaffles wrote: post #37
coldtea wrote: post #35

Earlier quoted context omitted.

I used ChatGPT. Isn't it already based on GPT-4 since a few weeks ago? It's the "Mar 23" version.

Unless you're paying for plus and then select the gpt-4 model, it's not 4. alternatively, you can sign up/request for api access here - https://openai.com/waitlist/gpt-4-api

All GPT-4 output should include a blue checkmark, so everyone can tell you payed for a subscription, and block you. ;)
Post reply on HN