36 comments

[ 2.5 ms ] story [ 80.7 ms ] thread
Go to the actual audio samples page to hear several interesting examples: https://google.github.io/tacotron/publications/tacotron2/ind...
Hearing the tongue twisters pronounced perfectly is a joy.
Looks like we found a new Turing test.
Wait until we train speech bots to fumble at tongue twisters

Uncanny valley

I wonder if it's more difficult to train something like this to stumble on tongue twisters in the same ways and places that a real person would.

Edit: Interesting, "Talib Kweli confirmed to AllHipHop that he will be releasing an album in the next year." actually sounds flawed to me. I hear "Al Hip Hop" instead of "All Hip Hop," though given my hearing problems, I could be wrong.

I dislike Google but damn if this isn’t impressive!
Maybe I start with the opposite prior, but this is absolutely mind blowing to me. The future truly looks much brighter now. If they couple this with their automatic translation (which will only get better), in a few short years you will be able to travel to an arbitrary country and use your smart phone to chat with people. Separately, I dream of the day when someone will release a personal language coach. I used one to improve my English, and I paid more than $xx per hour (a lot). Can you imagine paying $2.99 to get one with no time limits?
Consider checking out Italki.com

Personalized language teachers at extremely affordable rates. I've payed as low as $6/hr before.

I wonder how good the security/data-sanitation is. If you don’t know the language, you might think you’re saying “I will not buy this record, it is scratched.” but you end up with hovercraft eel removal services.
I think I'd still pick out most of these as likely machine-generated. But it's getting very close. Some of Amazon's Polly voices (especially a couple of the English women) sound good to me as well. They're not quite at the point where I'd use them for applications where people expect human speech but it's not far away.
I use Polly for ebook to audiobook conversion and they work pretty well for most stuff - not completely natural by any means but more than sufficient for casual listening.

https://auditus.cc

I had no trouble picking out the pairs at the bottom of that site (whether they are generated or not is in the URL for the audio, so you can do a blind test fairly easily)
The last set of samples includes comparisons with the source human. The only one that I can guess with confidence is the last pair. (Sample 1 of the pair contains assumed-context emotional inflection.)
I find the machine-generated ones to be more rhythmic and to have intonations with more consistent rises and falls in tone. I guessed the last three correctly based on listening for this. The first one I guessed wrong but in that case I was trying to differentiate them based on other tells.
Agreed, Polly has some great voices as well.
Just wait until they start using the voice data they've been collecting from millions of people via "OK <brandname>" enabled devices to imitate individual consumers' voices and use them to endorse products/brands to their contacts.
"However, the system is only trained to mimic the one female voice; to speak like a male or different female, Google would need to train the system again."

If I understand this correctly, they trained it to mimic a specific person, not just a human voice. That's... kind of terrifying. Trained to my voice, and given the right text, you could probably get me fired, possibly divorced, and maybe jailed. Trained to mimic Trump's voice, and given the right text, you might be able to start a war.

In the next ten years, I think there's going to be plausible deniability that any person's presence in a video, and especially in audio, is actually representative of them. There's already imperfect transpositions of celebrity faces onto the bodies of other people.

I imagine it's going to have a big impact when it first gets rolled out to mimic someone on the political stage, and then people will use it to discount any audio or video they don't like.

> ... and then people will use it to discount any audio or video they don't like.

I consider that to also be a big impact. If recorded events are now only hearsay, it's going to be a lot harder to prove what actually happened - both in politics and in court.

Although let's remember that it was that way for the vast majority of human history. It's only in the last 150-200 years that we've had light and sound recording technology.
It pretty terrifying in politics when in many countries like the US what happened usually matters less than the political affiliation of the messenger.
Don't courts rely on a chain of custody?
> We present a novel approach for real-time facial reenactment of a monocular target video sequence (e.g., Youtube video). The source sequence is also a monocular video stream, captured live with a commodity webcam.

Wow, this is impressive. Thank you for sharing! If that made its way on Hacker News on its own already, it surely deserved it.

For those who didn't go to the link, there's a short video demonstrating the technology -- one example shown is a YouTube video of George W. Bush being modified real-time by a webcam stream reenacting the facial expressions of the source user superimposed over the target.

(comment deleted)
I know nothing in text to speech, so maybe my question is stupid, but I've always wondered if somebody tried to produce "natural" sound by modelizing the air through a human mouth+throat+nose, so that your would have the "naturalness" of the voice (especially if you add the dynamics part, like air volume in the lungs that force you to pause, time it takes to reposition the tongue/mouth between two sounds) , or if it's actually more complicated than that/too ressource heavy/too hard to modelize etc.
I think the point of machine learning is that such detailed models do not need to be explicitly modeled, but as some implicit layers to learn.
It's better to have the right explicit features in your model than implicit ones. It's just harder to know how to make those features, and to know that they're right, which is part of the appeal of implicit features.
What is it they do to voices in audio recordings that make them sound like this?

I can’t put my finger on it but like all Audio Books have this distinctly processed sound. Is it some kind of voice auto tune or?