Randy Travis got his voice back with AI. The harder question is who else had to say yes

On 3 May 2024, Randy Travis shared "Where That Came From" with fans, his first new music in more than a decade. Travis can't sing it. In July 2013 he was hospitalised with viral cardiomyopathy and then had a stroke, and aphasia still limits his speech, the Associated Press reported. The voice on the record is his, generated by AI from his old recordings and laid over another singer's performance.

On 9 April 2025, Travis was in Washington as senators reintroduced the NO FAKES Act, and his wife Mary read a statement in his name against AI that takes a voice without permission, the Tennessean reported. The same technology, two opposite verdicts. The difference between them is the whole argument, and it's worth being exact about.

How was the vocal actually made?

According to the AP's account:

• Cris Lacy of Warner Music Nashville brought the idea to the Travises and had developers in London build a proprietary model.
• The team ended up with two models: one trained on 12 vocal stems, the other on 42 stems from across Travis's career, 1985 to 2013. CBS described the 42 as tracks pulled from the label's vault with the music stripped out.
• The song, by Scotty Emerick and John Scott Sherrill, was one Travis's longtime producer Kyle Lehning had co-produced and held on to for years. The demo vocal was sung by James Dupré (AP spells it Dupree), a Louisiana singer who has been a Travis stand-in and tours with his original band, according to KTDY.
• Lehning said the model took about five minutes to process the demo, and the raw result was "probably 70, 75% what you hear now". Lehning and engineer Casey Wood then edited it by hand, adjusting things like vibrato speed and phrase timing.

Lehning told CBS that having Travis there as "a vital part of the decision-making process makes all the difference to me." That matters for what kind of AI this is. It's voice conversion, not text-to-song. Melody, phrasing, breath and timing start as a human performance; the model supplies the timbre; people fix what doesn't sound like Travis.

What did the Travises say about deepfakes?

In 2024, Mary Travis told the AP the team was "hoping that maybe we can set a standard", and argued that credit should go where it's due and that artists should control their voices. A year later, the statement she read for Randy in Washington drew the line explicitly. The recordings, it said, "were produced as an extension of my artistry" and are "very different" from "someone else stealing my voice and producing music." It ended: "It is unconscionable that AI can take my voice and produce content without my consent or compensation."

The bill they backed, S. 1367, was introduced that day by Sen. Chris Coons with Sens. Marsha Blackburn, Amy Klobuchar and Thom Tillis. As introduced, it would:

• create a property right in a person's voice and visual likeness that can't be assigned during their life but can be licensed;
• cap an adult's licence at 10 years, in writing, signed, and describing the intended uses;
• run the right for 10 years after death, renewable in five-year blocks while the voice stays in authorised public use, up to 70 years;
• make online services take down replicas on notice, with statutory damages starting at $5,000 per work for an individual and $25,000 per work for other entities.

The Tennessean reported that the bill builds on Tennessee's ELVIS Act. This wasn't the couple's first trip to the Hill: in June 2024 Mary delivered Randy's remarks to a House Judiciary subcommittee on radio royalties, per SoundExchange.

Did everyone underneath say yes?

The story is usually told as a consent success, and for Travis it was. For the singer under the vocal, the record is murkier. Wikipedia's article on the song, citing Rolling Stone, says Dupré recorded the vocal in 2011, hadn't given permission for this use, learned of it on the day of his CBS interview, and then gave permission after the fact and was paid. We couldn't open the Rolling Stone piece, and the AP and CBS reports we read don't address it, so treat that as reported rather than confirmed. CBS called his voice "donated".

Even on the most generous reading, it shows a gap. Under S. 1367 a digital replica is a realistic representation "readily identifiable" as a person's voice. Dupré's timbre is gone from the finished track, so as we read the bill, his contribution wouldn't be a replica of him at all. What the law protects is the voice you hear, not the performance underneath it.

What we think

"Where That Came From" is the best case anyone has made for AI voice work, and it's persuasive precisely because it's narrow. The model was trained on one artist's own recordings, for one song, with the artist, his family and a producer of decades present, and people making the final calls. That's the opposite of a general-purpose voice model anyone can prompt. The useful distinction isn't "AI or not". It's whether the artist directs the tool or the tool replaces the artist.

Three caveats keep it from becoming a template others can copy. First, it ran through the label: the stems came from Warner's vault, so "artist-controlled" depended on a label that wanted to cooperate. An artist whose masters sit with an indifferent owner, or who has no clean isolated vocals, can't do this. Second, Lacy told CBS, "It's still his vocal." In timbre, yes. In performance, it's partly Dupré's, and the credit should say so. Third, the consent that the NO FAKES debate focuses on is the star's. Surrogate singers, session vocalists and demo singers need their own: a contract before their take is used, pay, and a credit. The bill's 10-year licence cap and ban on assignment during life are good protections for the named artist. Nothing equivalent covers the person whose phrasing the model inherits.

The principle the Travises stated, AI as an extension of the artist rather than a stand-in, is the right test for production tools generally, and it's the one browser tools such as Veena, which describes an AI co-producer that builds tracks with you "layer by layer" and says everything you make stays yours, should be judged by too. Labels that want the goodwill this song earned should copy the whole package, including the singer underneath.

Sources

• Associated Press (Maria Sherman), With help from AI, Randy Travis got his voice back (6 May 2024); read via ClickOnDetroit's syndicated copy
• CBS News, More than a decade after a stroke, Randy Travis sings again, courtesy of AI (5 May 2024)
• MusicRow, Making Of Randy Travis' AI-Generated Single Explored In CBS News Sunday Morning Segment (3 May 2024)
• 99.9 KTDY, Louisiana Singer is the 'Voice' Behind New Randy Travis Song (6 May 2024)
• Wikipedia, Where That Came From
• The Tennessean (Melonee Hurt), Randy Travis speaks out against AI-enabled deepfakes (9 Apr 2025); read via Yahoo News
• GovInfo, S. 1367, NO FAKES Act of 2025, as introduced (9 Apr 2025)
• Recording Academy, NO FAKES Act
• SoundExchange, Huppe and Randy Travis call on Congress to approve the American Music Fairness Act (27 Jun 2024)
• Veena Studio, Veena