Deepfake Voice Technology: The Good. The Bad. The Future

https://www.econotimes.com/Deepfake-Voice-Technology-The-Good-The-Bad-The-Future-1601278

by respeecher • 5 years ago

85 35 5 years ago

falcolas• 5 years ago

There’s a whole other avenue that’s not covered here: theft of voice actors voices. It was covered by a voice actor on YouTube, as a warning for other actors. The first is only two minutes or so, the second is a bit over ten, but goes into more detail.

https://youtu.be/p5onpLkhq50

https://youtu.be/vr7YbfzSd5A

onethought4• 5 years ago

It's an interesting question, copyright doesn't cover tone/pitch/etc... so how do you protect that? Should you even be allowed to?

dave_sullivan2• 5 years ago

As someone that wants to abolish patents and limit copyright, I'd say no. Others without those leanings will no doubt disagree.

tyingq2• 5 years ago

Bette Midler won her lawsuit against Ford Motor Co: https://en.wikipedia.org/wiki/Midler_v._Ford_Motor_Co.

lloda• 5 years ago

There's probably a trademark / misrepresentation / self image rights angle that could be argued, same as with say faces. That wouldn't seem unfair to me. It should require a lot more than just pitch similarity imo.

jonathanstrange• 5 years ago

It may be a grey zone now but I find it likely that judges will decide that copyright does cover the sound of someone's voice and other aspects (intonation, accent), as these are vital aspect of the performance and often the reason why a voice actor gets hired.

I'd go much farther than that. Unique features of someone's voice and appearance (face, walk) should not only be protected by copyright but also by privacy laws. It would be crazy if producers were allowed to make arbitrary videos with anyone's appearance and voice, putting words in your mouth you'd never endorse.

Luckily, this may be one of the cases when politics will react fast and swiftly, if necessary with new laws, since politicians would likely be among the first victims of the new trend.

skocznymroczny• 5 years ago

Polish game modders for the game Gothic (very popular RPG in Poland) are using "AI" to reproduce the voices of the Polish voice actors who did the original dub of the game. The quality isn't there yet, the voice clarity is all over the place, but it's already a much better option than using a generic "Microsoft Sam" kind of voice. And when technology matures, it will allow modders to have in-game characters sound as they "should sound like", instead of being silent or resorting to amateur dubbing as it is now.

modeless• 5 years ago

I have yet to hear any example of good quality voice deepfakes. Every example I've heard so far was completely unconvincing, and absolute trash compared to human impersonators. Speaking of, human voice impersonation has been possible and quite convincing for forever.

Audio deepfakes, when they finally arrive for real, will simply reduce the effort required for voice impersonation. They do not make possible new feats of impersonation that were never possible before. They can still have an effect on society for sure, but I believe the concern is out of proportion with the effect.

chrisco255• 5 years ago

The effort will be reduced to typing sentences and selecting emotions from a dropdown menu. That opens up a whole new can of worms for a society that already has trust issues.

exporectomy2• 5 years ago

Is that your own society? How would it exacerbate your trust issues? Surely making people doubt untrustworthy things is good?

BikiniPrince• 5 years ago

The Jordan Peterson df was quite good. I made a few sound bytes to send to some friends as a gag, but they were already aware it was in the wild. The compression is probably the worst here, but it was fun when it was up. https://www.notjordanpeterson.com/

spyder• 5 years ago

It was already successfully used to "Scam A CEO Out Of $243,000":

https://www.forbes.com/sites/jessedamiani/2019/09/03/a-voice...

Probably it' easier because phone calls aren't high quality.

arnaudsm• 5 years ago

How did they prove the voice was an AI and not a human impersonator ?

ansk• 5 years ago

If you look at the latest research in neural network based approaches to text-to-speech, you'll find plenty of examples of realistic synthetic voices. However, due to the nature of the available training data (audiobooks), these models 1) are restricted to reading text in a narrative style and 2) aren't generally trained on speech of recognizable public figures. There was a paper a while back that was trained on TED talks and produced fairly convincing voices, some of which you may recognize (Bill gates, Stephen Wolfram and some others - see here https://sjvasquez.github.io/blog/melnet/) Definitely not 100% convincing, but I wouldn't call it absolute trash.

egfx• 5 years ago

Check out the demos at https://www.descript.com/overdub

I wouldn’t call the voices unconvincing. They used to have a better demo though where it would impersonate celebrities like Trump. It was very interesting to demo it.

dbspin• 5 years ago

Obviously it's subjective, but all of these demos have the 'watery' digital sound that's an immediate giveaway. By definition they're unconvincing.

kenoph• 5 years ago

I am using descript for a project of mine. You can definitely still hear that they are AI-generated. But I'm sure it's not gonna take long. Couple this with SSML and it's going to be quite convincing.

bryanrasmussen• 5 years ago

automating sending out fake emails allowed any idiot to do it.

there is a difference in scale and effect between every idiot in the world can fake as many impersonations of people that they want (and they can coordinate this impersonation with 100s of others who have a goal with this impersonation) and Rich Little calling up somebody's granny and claiming to be Jimmy Stewart.

sillysaurusx• 5 years ago

I made Dr Kleiner sing "I am the very model of a modern major general": https://www.youtube.com/watch?v=koU3L7WBz_s&ab_channel=Shawn...

It's pretty trash, but it's not completely horrible. It's also over a year old, so the SOTA is likely better today.

chrisin2d• 5 years ago

They can still have an effect on society for sure, but I believe the concern is out of proportion with the effect.

I believe the contrary: that high-quality, convincing audio (and video) deepfakes will be cataclysmic for society. Voice audio is one of the last digital artifacts that people will believe with their senses.

When Donald Trump’s “grab them by the p—“ audioclip came out, he was at least forced to acknowledge it.

Once high-quality audio deepfake tools become available, amateurs will be able to flood the web with fake audioclips of politicians saying things they never did, and truly unscrupulous politicians will be given technological cover for any damning audioclips of theirs that get leaked.

Once we cannot trust our senses, then we will be only able to trust our institutions and then our tribes. But trust in institutions is already decaying rapidly, so in the end there only ideological tribes will be left.

SubiculumCode2• 5 years ago

Do not underestimate the power of mediocre at scale.

Synthetic is scalable.

gotrythis• 5 years ago

Did you check out the company linked in the article?

https://www.respeecher.com/

PeterStuer• 5 years ago

This is an infomercial, by Respeecher.

bitlevel• 5 years ago

This guy speaking - with Obama's voice - is eerie.

https://www.youtube.com/watch?v=t5yw5cR79VA&feature=emb_logo

illwrks• 5 years ago

I've had quite an odd idea this past year as video conferencing has taken off massively both for public figures and private individuals.

....what if you had a malicious video conferencing service that was free to all?

Because it's free it gains mass adoption,the service can target a person it is interested in and can then record all audio and video footage.

Then using this captured data train a real-time puppet.

You could then intercept any call that person is on and 'take over' and push your own agenda, or call people up and do the same, pretending to be that person.

With something like that it would add more value in ongoing OTP ways of verifying a person is who they say they are.

SubiculumCode2• 5 years ago

Another good reason not to keep research and development of companies like Zoom in America.

FloatArtifact• 5 years ago

I have a feeling the text to speech market is suppressed. For text-to-speech services such as Google or Amazon it costs more (end user perspective) to process audiobooks from eBooks. Compared to buy the audiobook through a subscription service.

Deep fakes audio sounds better than most text to speech models that are freely available.

Deepfake Voice Technology: The Good. The Bad. The Future

Related Stories