ZeroSpeech Challenge 2019: TTS without T | Hacker News Reader

ZeroSpeech Challenge 2019: TTS without T

zerospeech.com

47 pointsivansavz7 comments

The final paper submissions deadline was April 5 2019, so we can check the results now:

https://zerospeech.com/2019/results.html

^ click on the green (+) to see details for each submission including audio samples reconstructed.

Super interesting. Yet IMO kids have more than a text available. They have several senses, most importantly their eyesight. Hence this comparison does not seem to be on spot.

But blind kids also learn quickly to talk

They have touch, smell and taste.

Evil much? This is, as described, a way to impersonate anyone else on the phone. I can see the abuse of that capability combined with caller ID spoofing, a crook could drain the bank accounts of half the elderly population of any developed country.

so could anyone with a predilection towards voice acting and a proper understanding of how to imitate others. one way or another, bad things are going to happen because there will always be people who will use their skills or tools for purposes other than what they are intended for. i don't think it's a valid argument to dismiss the utility and power of something just because people might do what people have always done.

I see it as the difference between APTs and script kiddies. Anyone with an investment of time, and a certain minimum of skill, can reconnoiter a system, identify a weakness, exploit it for access and do bad things. The level of investment is high, so the expected reward has to be worth it. Once exploits are packaged into a piece of code you can just install and run and have it break into random systems, the investment is low so you for systems vulnerable to those attacks you get many many more incidents.

Finding the requisite voice acting skills, training for the voice in question, and then identifying a number and calling it is a lot of investment in time and money and resources. (voice actors aren't cheap, especially if you're asking them to defraud someone's grandparents)

But if this project is successful in being able to mimic a person's voice saying what ever you want it to say, then anyone hanging around a theme park could capture voice signatures, and relations, and then turn around and exploit the same for fraud without anyone else getting involved.

To me, that seems like it would open up a large attack surface that is currently poorly defended at best.

Reply on news.ycombinator.com