Part 1 of 3 in our PlaiPals series.
Your voice is the one instrument nobody else can buy at a music store. You can copy a guitar tone. You can buy the same synth as your favorite producer. But the sound that comes out of your throat has always been yours.
For most of recording history, that was simply a technical fact. Machines could make a voice louder, higher, stranger or more robotic. They could not make you sound like somebody else.
That changed fast. Here's how it happened, and why the missing ingredient was never the technology. It was permission.
Before AI: tools that changed the sound, not the singer
Start with the vocoder. In the mid-1930s, a Bell Labs engineer named Homer Dudley built it to squeeze speech into less bandwidth so it could travel more cheaply over long-distance lines (MIT Technology Review). It broke a voice down into simple signals and rebuilt it, which is why it sounds so robotic. That robotic quality is what later made it a studio effect. But a vocoded voice is still the voice that went in.
Then came Auto-Tune. Antares introduced it in 1997 as a quiet fix for off-pitch notes. A year later, Cher's "Believe" turned it into a deliberate, glitchy effect that the whole world could hear (Britannica). Auto-Tune changes pitch. It can make a note snap or warble. It does not change who is singing.
Pitch shifters and harmonizers work the same way. Push them too far and you get chipmunks or monsters, not a different human being.
Singing synths: built with singers, sold as characters
Yamaha's Vocaloid is the closest pre-AI tool to "a voice in a box." The first Vocaloid voice libraries went on sale in 2004, and Hatsune Miku arrived in 2007 and became a global character (Vocaloid official history).
Vocaloid voices were made from recordings of real people, but they were sold as their own characters, not as copies of famous singers. When a real star's voice did become a Vocaloid, it was sold openly under his name. Vocaloid's own history describes Gackpoid, a product designed to "faithfully reproduce the voice" of Japanese singer and actor Gackt (Vocaloid official history).
In other words, the voice-synth world built its voices out in the open, as named products. Nobody was quietly passing off a copy of a star.
The only way to "sound like" someone: hire a human
Before AI, the only way to get a famous voice you couldn't book was to hire a soundalike. Courts did not love that.
In 1988, a federal appeals court ruled in Midler v. Ford that deliberately imitating a well-known singer's distinctive voice to sell a product could be a violation of that singer's rights, even though no recording of hers was used (Justia). In 1990, Tom Waits won $2.5 million over a snack ad that used a singer imitating his voice (Vulture).
The lesson was already on the books decades before AI: a voice is part of a person's identity, and using it to trade on their name without asking can cost you.
The licensed way: real singers, real permission
Producers who wanted real vocals on a track had a clean option: sample packs and loops. On platforms like Splice, a singer records phrases, those sounds are licensed, and producers can use them in commercial releases royalty-free (Splice Licensing FAQ).
Look closely at the rules, though. Splice's license says you can't use the name or likeness of the creator behind a sound, which can include their voice, without their express permission. It also says its sounds can't be used as AI training data (Splice Licensing FAQ).
So the licensed world drew a clear line. You can use the recording the singer chose to share. You can't turn the singer into your featured artist.
Around 2020 to 2023: the machines learn to impersonate
Around 2020, AI experiments trained on famous artists' music without their consent were already raising eyebrows. Holly Herndon, for example, spoke publicly about the ethics of "TravisBot," an AI trained on Travis Scott's music without his consent (The FADER).
Then, in early 2023, the floodgates opened. A free, open-source tool called so-vits-svc (short for SoftVC VITS Singing Voice Conversion) made "voice conversion" possible on a decent home computer (so-vits-svc on GitHub). Voice conversion means you sing or rap a part, and the software re-renders it in someone else's voice.
As Vulture reported in May 2023, you could train a model on 20 to 30 minutes of clean a cappella audio, or just download a model someone else had made. Shared models of famous pop stars were easy to find, and YouTube and TikTok filled up with AI covers credited to famous voices (Vulture). A second open-source project, RVC (Retrieval-based Voice Conversion), appeared that spring with the pitch that you could train a good model on 10 minutes of voice data or less (RVC on GitHub).
Here's the uncomfortable part. The most popular early voice models were copies of famous singers, and from what was publicly reported, those singers were not asked. Their voices were pulled from their own records and passed around as files.
April 2023: "Heart on My Sleeve"
In April 2023, an anonymous TikTok user called Ghostwriter977 released "Heart on My Sleeve," a song with AI vocals that sounded strikingly like Drake and The Weeknd (Billboard).
It went viral in days. By the time it came down on April 17, it had about 600,000 Spotify streams, 15 million TikTok views and 275,000 YouTube views, according to The Guardian (The Guardian). Universal Music Group, the artists' label, condemned "infringing content created with generative AI." Apple Music, Deezer, TIDAL and Spotify pulled the track (Billboard), and YouTube's notice said the video was no longer available "due to a copyright claim by Universal Music Group" (Variety).
UMG also asked the industry "which side of history" it wanted to be on: "the side of artists, fans and human creative expression, or on the side of deep fakes, fraud and denying artists their due compensation" (Vulture).
Notice what happened. The real issue was a voice, but the fastest tool available was a copyright claim. In 2023 there was no simple national rule that said "you can't clone a singer."
2024: the labels go to court
The next fight was bigger than voice cloning. It was about how AI music generators were built.
On June 24, 2024, record companies including Sony Music, Universal Music Group and Warner Records sued two AI song apps for copyright infringement. The labels accused them of copying recordings without permission to train systems that can spit out full songs from a text prompt. One case was filed in federal court in Massachusetts and the other in New York (AP News; Court filing, via RIAA).
Those cases were mainly about recordings used for training, not cloning one singer. But the root problem was the same: artists' work used to build AI products without asking.
The story didn't end there. Some of those labels later settled and signed licensing deals, which we cover in Part 2. Others are still in court. As of September 2026, Sony and Universal are still suing one of those apps, and a group of musicians including Jason Isbell has filed a proposed class action claiming that app lets users generate songs that evoke specific artists' identities without consent. That company disputes those claims (Los Angeles Times).
What this history tells us
We love new tools at PLAIGROUND. Auto-Tune, vocoders, synths and samples all became part of how music is made.
AI voice tools can go the same way. The problem with the first wave wasn't that the tech was new. It was that singers were treated as raw material instead of collaborators. Nobody asked. Nobody credited them. Nobody paid them.
That's fixable. In Part 2, we'll look at the artists, companies and lawmakers who started fixing it, and why a singer's voice is now being treated as its own right, separate from the song and the recording.
Sources
- MIT Technology Review, "Vocal Codes" (Nov. 1, 2003): https://www.technologyreview.com/2003/11/01/274977/vocal-codes/
- Britannica, "Auto-Tune": https://www.britannica.com/technology/Auto-Tune
- Vocaloid (Yamaha), "20 years of VOCALOID history": https://www.vocaloid.com/en/anniversary/history
- Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988), via Justia: https://law.justia.com/cases/federal/appellate-courts/F2/849/460/37485/
- Vulture, "AI Singers Are Unnervingly Good and Already Ubiquitous" (May 1, 2023): https://www.vulture.com/article/ai-singers-drake-the-weeknd-voice-clones.html
- Splice, "Splice Sounds Licensing FAQ": https://support.splice.com/en/articles/8652642-splice-sounds-licensing-faq
- The FADER, "Holly Herndon announces new voice instrument and 'digital twin' Holly+" (July 14, 2021): https://www.thefader.com/2021/07/14/holly-herndon-holly-plus
- so-vits-svc (SoftVC VITS Singing Voice Conversion), GitHub: https://github.com/svc-develop-team/so-vits-svc
- RVC (Retrieval-based Voice Conversion WebUI), GitHub: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI
- Billboard, "Fake Drake & The Weeknd Song, Made With AI, Pulled From Streaming After Going Viral" (Apr. 17, 2023): https://www.billboard.com/pro/fake-ai-drake-the-weeknd-song-pulled-streaming/
- Variety, "AI-Generated Fake 'Drake'/'Weeknd' Collaboration, 'Heart on My Sleeve,' Delights Fans and Sets Off Industry Alarm Bells" (Apr. 2023): https://variety.com/2023/music/news/fake-ai-generated-drake-weeknd-collaboration-heart-on-my-sleeve-1235585451/
- The Guardian, "AI song featuring fake Drake and Weeknd vocals pulled from streaming services" (Apr. 18, 2023): https://www.theguardian.com/music/2023/apr/18/ai-song-featuring-fake-drake-and-weeknd-vocals-pulled-from-streaming-services
- AP News (June 24, 2024): AP News
- Court filing (D. Mass., filed June 24, 2024), via RIAA: Court filing
- Los Angeles Times, "Why some musicians aren't happy about labels signing AI deals with platforms" (Sept. 21, 2026): https://www.latimes.com/entertainment-arts/business/story/2026-09-21/ai-music-licensing-deals-labels-artists-who-benefits