Randy Travis ‘Gets Voice Back’ Using AI
A technological miracle with obvious potential for abuse.
AP (“With help from AI, Randy Travis got his voice back. Here’s how his first song post-stroke came to be“):
With some help from artificial intelligence, country music star Randy Travis, celebrated for his timeless hits like “Forever and Ever, Amen” and “I Told You So,” has his voice back.
In July 2013, Travis was hospitalized with viral cardiomyopathy, a virus that attacks the heart, and later suffered a stroke. The Country Music Hall of Famer had to relearn how to walk, spell and read in the years that followed. A condition called aphasia limits his ability to speak — it’s why his wife Mary Travis assists him in interviews. It’s also why he hasn’t released new music in over a decade, until now.
“Where That Came From,” which released Friday, is a rich acoustic ballad amplified by Travis’ immediately recognizable, soulful vocal tone.
Cris Lacy, Warner Music Nashville co-president, approached Randy and Mary Travis and asked: “‘What if we could take Randy’s voice and recreate it using AI?,’” Mary Travis told The Associated Press over Zoom last week, Randy smiling in agreement right next to her. “Well, we were all over that, so we were so excited.”
“All I ever wanted since the day of a stroke was to hear that voice again.”
Lacy tapped developers in London to create a proprietary AI model to begin the process. The result was two models: One with 12 vocal stems (or song samples), and another with 42 stems collected across Travis’ career — from 1985 to 2013, says Kyle Lehning, Travis’ longtime producer. Lacy and Lehning chose to use “Where That Came From,” a song written by Scotty Emerick and John Scott Sherrill that Lehning co-produced and held on to for years. He believed it could best articulate the humanity of Travis’ idiosyncratic vocal style.
“I never even thought about another song,” Lehning said.
Once he input the demo vocal (sung by James Dupree) into the AI models, “it took about five minutes to analyze,” says Lehning. “I really wish somebody had been here with a camera because I was the first person to hear it. And it was stunning, to me, how good it was sort of right off the bat. It’s hard to put an equation around it, but it was probably 70, 75% what you hear now.”
“There were certain aspects of it that were not authentic to Randy’s performance,” he said, so he began to edit and build on the recording with engineer Casey Wood, who also worked closely with Travis over a few decades.
The pair cherrypicked from the two models, and made alterations to things like vibrato speed, or slowing and relaxing phrases. “Randy is a laid-back singer,” Lehning says. “Randy, in my opinion, had an old soul quality to his voice. That’s one of the things that made him unique, but also, somehow familiar.”
His vocal performance on “Where That Came From” had to reflect that fact.
“We were able to just improve on it,” Lehning says of the AI recording. “It was emotional, and it’s still emotional.”
Mary Travis says the “human element,” and “the people that are involved” in this project, separate it from more nefarious uses of AI in music.
“Randy, I remember watching him when he first heard the song after it was completed. It was beautiful because at first, he was surprised, and then he was very pensive, and he was listening and studying,” she said. “And then he put his head down and his eyes were a little watery. I think he went through every emotion there was, in those three minutes of just hearing his voice again.”
Here’s the official video:
And another (better, in my judgment) version showing Travis reacting to his “own” voice:
My assessment of the result is pretty much the same as The Verge‘s Wes Davis (“Randy Travis gets his voice back in a new Warner AI music experiment“) adds:
The result of Warner’s experiment is a gentle tune that captures Travis’ relaxed style, which rarely wavered far from its baritone foundation. It sounds like one of those singles that would’ve hung around the charts long enough for me to nervously sway to once after working up the gumption to ask a girl to dance at a middle school social. I wouldn’t say it’s a great Randy Travis song, but it’s certainly not the worst — I’d even say I like it.
It’s on par with, say, the recent new Rolling Stones singles: immediately identifiable, a perfectly worthy addition to their catalog, and yet not as good as their work in their prime.
As to the implications, I’m on a similar page there as well:
Dustin Ballard, who runs the various incarnations of the There I Ruined It social media account, creates his AI voice parodies in much the same way as Travis’ team, giving birth to goofy mash-ups like AI Elvis Presley singing “Baby Got Back” or synthetic Johnny Cash singing “Barbie Girl.”
It would be easy to sound the alarm over this song or Ballard’s creations, declaring the death of human-made music as we know it. But I’d say it does quite the opposite, reinforcing what tools like an AI voice clone can do in the right hands. Whether you like the song or not, you have to admit that you can’t get something like this from casual prompting.
Cris Lacy, Co-president of Warner Music Nashville, told CBS Sunday Morning that AI voice cloning sites produce approximations of artists like Travis that don’t “sound real, because it’s not.” She called the label’s use of AI to clone Travis’ voice “AI for good.”
Right now, Warner can’t really do much about AI clones that it feels don’t fall under the heading of “AI for good.” But Tennessee’s recently-passed ELVIS Act, which goes into effect on July 1st, would allow labels to take legal action against those using software to recreate an artists’ voice without permission.
Travis’ song is a good edge-case example of AI being used to make music that actually feels legitimate. But on the other hand, it also may open a new path for Warner, which owns the rights to vast catalogs of music from famous, dead artists that are ripe for digital resurrection and, if they want to go there, potential profit. As heartwarming as this story is, it makes me wonder what lessons Warner Music Nashville — and the record industry as a whole — will take away from this song.
I’m more leery than optimistic about how this technology will be used in the future. I’m actually fine with the mashup trend but really don’t want new “Elvis” or “John Lennon” songs that are put out by record companies.
But, yes, this is a case of “AI for good” because Travis himself had final say over the project. While it doesn’t, despite the universal headlines to the contrary, actually give him his singing voice back, it’s the next best thing. So long as he’s happy with the result and fans are clear as to what they’re getting, there’s no down side to a new Randy Travis release that brings a smile to his face.
I’m glad for RT as a person and artist – being trapped inside your own body after a stroke is hell on earth. But I think society as a whole needs to start seriously considering how AI has the potential to mess up our lives. This would not be the first time we’ve embraced new technology without seeing the big picture.
@Not the IT Dept.:
Shhhhh, no, no, no, our tech bro betters have assured us that like trickle down economics, this will make life better for all of us…
That likely started when it was discovered that a hefty club was not only useful for killing prey for food but also killing anyone who might want to compete for your meal.
@Beth: I would note that there are things that it is impossible to not do.
Technology develops along a broad path and at some point you note that “X is possible”. You may not think X is a good idea to do, but someone will, and they will do it. Usually because they can make money doing it, but also because it’s cool.
There’s really no stopping this kind of thing. Well, you can do this for a while if you have an ultra authoritarian state, (I’m thinking of old Imperial China – Ming Dynasty, etc.) but doing that kind of thing as a regular thing leaves you way behind after a while.
@Jay L Gischer:
I agree with you that it’s probably impossible to stop most if any technological progress. That being said, there is a big difference between not trying to stop something and diving in with wild, absurd, reckless abandon. It seems to me that the techbro class (and they are almost universally men) and the tech finance class (slightly more non-men) have decided that if they can do something, they should have absolutely unfettered ability to do so and that everyone else is responsible for whatever happens.
The second part is what bothers me to no end. I’m sure there a good, legitimate, positive uses for AI, but we’re being fed crap and being told we’re idiots for not wanting to eat it.
I don’t see how this is any more nefarious than auto-tune. That technology has already ruined vocal performance in commercial music.
Intellectual property rights are certainly going to be an issue. If my AI program generates a Elvis sound alike song, who gets the money? John Fogerty got sued, unsuccessfully, for putting out a song that sounded like a CCR song.
@Beth: Well, I’m not thrilled about your dismissal of the “techbro class”. It’s true that about 85 to 90 percent of the programmers I’ve worked with are male. It’s also true that most of those people (male or female) actively made choices about the impact of stuff they were willing to work on.
The point I want to make is the broad brush class you reference is so very, very vague it can easily erase all those good people I worked with. Thing is, there’s always someone. There’s always that person who does stuff that means we can’t have nice things. This is a phenomenon that is much, much broader than tech.
We notice tech now because it’s very powerful. When I started here in the 80’s, nobody took much notice of us, because we weren’t so powerful.
To me, the important thing to note about an AI is whose interests does it serve? The Algorithm (of YouTube to be specific) serves the interests of YouTube first and foremost. It does that in a way that is also somewhat helpful to me, but can easily turn addictive and harmful.
And that’s true of the OP. In this case, the AI served Randy Travis’ interest first and foremost, and in a way that does little harm to anyone else. So doubleplus good! Of course, not so good uses of it are easy to imagine. Very easy.
@Slugger: My non-lawyer take is that if you claim that the recording is Elvis, or promote it as “sounds like Elvis” you are liable. But otherwise it is no different than a person doing their best Elvis impersonation without ever mentioning Elvis, which from what I can tell, is legal.
Just wanted to chime in to signal boost team “AI for good.” My group (and collaborators) have been working in this space for many years now. Since before LMs were LLMs. These and other advances have accelerated our work and opened more new better opportunities to use AI for good.
Our use cases have mostly focused on understanding clinical decision-making and, subsequently, using that knowledge to help patients and clinicians more effectively partner together. Much of this work has focused on really difficult health conditions/situations. And with marginalized and under-resourced communities.
I say this mostly to brag… ahem… to reinforce that there is a lot of good being done (and attempted). For/with celebrities like Randy Travis. And also for/with the “invisible” folks who don’t get featured. Yes and.
@Jay L Gischer: Did Elvis Presley’s estate license his voice and recordings to the AI company that trained their model on it?
We don’t have case law yet that covers AI and fair use, nor has congress weighed in. How much of their voice was used by the training model? Is it transformative? Does it create confusion in the marketplace?
My favorite anti-AI people spend time trying to get AI to violate Disney copyrights. The goal is to pit Disney off against the techbros and venture capitalists, because in such a matchup they’re just rooting for blood.
@Gustopher: You are correct in that the law is not settled in this area at all. And given the commercial interests on either side are not small, I don’t expect it to get settled in a reasonable way for a long time.
AND, I think a firm using Elvis recordings, legally obtained, to train an AI could readily make the claim that what they are doing is no different from a person who develops a good Elvis impersonation by listening to Elvis records.
What it turns on legally, to my mind, is whether you are representing yourself or your work as Elvis.