AI singing voices are produced by models that turn a musical score into a spectrogram of sung audio, or that reshape a recording to sound like a chosen singer. The first approach is documented in research such as DiffSinger, accepted at the AAAI conference in 2022. The second, when it copies a real artist without permission, is what the U.S. Copyright Office calls an unauthorized digital replica, and as of October 27, 2026 the rules for it come from state law, with a federal bill still pending.
Key takeaways
- DiffSinger generates a mel-spectrogram of singing from a music score by starting from noise and refining it step by step.
- A song with AI-generated vocals imitating Drake and The Weeknd drew more than 15 million social media views in April 2023, although neither artist knew of it.
- Tennessee extended its right of publicity to cover voice simulations in a law that took effect July 1, 2024.
- The Copyright Office recommended a new federal digital replica law in a report dated July 31, 2024.
- The NO FAKES Act of 2026 advanced out of the Senate Judiciary Committee by voice vote on June 18, 2026. A bid to pass it by unanimous consent drew an objection on September 30, 2026, and it had not been enacted at the last check.
How a model sings from a score
Text-to-speech systems read words aloud, while singing synthesis must also follow pitch and note length. The DiffSinger paper, by Jinglin Liu and colleagues, first posted on May 6, 2021, describes an acoustic model for singing voice synthesis that conditions on a music score. It uses a diffusion probabilistic model: generation starts from noise and is iteratively turned into a mel-spectrogram, a spectrogram-style representation of the audio. The authors say diffusion training is more stable than training with generative adversarial networks and avoids the over-smoothing seen with simple L1 and L2 loss functions.
The paper also adds a "shallow diffusion mechanism" that starts partway through the process from the output of a simpler decoder, which the authors report improves voice quality and speeds up generation. Their evaluation used a Chinese singing dataset, and they report that the method outperformed earlier singing synthesis approaches. The same work was extended to speech under the name DiffSpeech. This describes one published research system. It does not describe how any particular commercial music product works internally, which vendors rarely document.
When the voice belongs to a real artist
The legal questions begin when a model is made to sound like a specific person. The Copyright Office's report on digital replicas opens with an April 2023 song known as the "Fake Drake" song, which drew over 15 million social media views and 600,000 Spotify listens. Neither artist was aware of it before release, because the vocals were unauthorized AI-generated replicas. The report notes that generative AI has become so sophisticated and accessible that minimal expertise is needed to produce such replicas.
The same report lists beneficial uses of digital replicas: accessibility tools for people with disabilities, "performances" by deceased or non-touring artists, support for creative work, and licensing deals through which individuals are compensated for the use of their voice, image, and likeness. Consent and control, rather than the technology itself, mark the line between those uses and the harmful ones.
Where the rules stand
The table separates what is law from what is a recommendation or a bill. It is a simplified orientation based on the sources listed below, not a complete legal survey.
| Measure | Status on October 27, 2026 | What it does or proposes |
|---|---|---|
| Tennessee ELVIS Act (2024) | Enacted; signed March 21, 2024, effective July 1, 2024 | Extends the state's right of publicity to voice simulations and to publishing, distributing, or making a voice available without permission, beyond advertising |
| Copyright Office report, Part 1: Digital Replicas | Published July 31, 2024; a recommendation, not law | Proposes a federal law covering all individuals, with liability for distribution rather than creation alone |
| NO FAKES Act of 2026 (S. 4591, H.R. 8915) | Pending; cleared Senate Judiciary Committee June 18, 2026; unanimous-consent attempt blocked in the Senate on September 30, 2026 (news reports) | Would create a federal right in a person's voice and likeness against unauthorized digital replicas |
The Copyright Office describes the Tennessee change as extending the publicity statute to cover voice simulations and to all acts of unauthorized publishing, performing, distributing, transmitting, or making available to the public. Its recommended federal design includes liability that requires actual knowledge that a replica was both of a particular person and unauthorized, a notice-based safe harbor for online service providers, licensing allowed but outright assignment not, and an express balance with free speech concerns. These are recommendations, and the 2026 bill, as described by its sponsors, adds free speech safeguards and exemptions for libraries, archives, and research institutions. Its final text could still change. Senator Ted Cruz objected on September 30, 2026, to passing it by unanimous consent, citing unclear protections for satire and parody, according to news reports, so floor action was still open at the last check.
Questions to ask about an AI-sung track
- Does the track claim to feature a real artist, and is there evidence the artist agreed?
- Is the voice a replica of a particular person, or a synthetic voice with no real counterpart?
- Who holds the rights to the underlying song, lyrics, and recording?
- Which jurisdiction applies? Rules differ between U.S. states and between countries.
- Does the platform where it appears have a process for reporting unauthorized replicas?
This article gives general information, not legal advice. Canadian law was not reviewed here, and anyone facing a specific dispute over a voice, a recording, or a license should consult a lawyer.
FAQ
Can AI really sing like a particular artist?
The Copyright Office report describes AI-generated vocals that can pass for a favorite artist, with the 2023 "Fake Drake" song as its example. Quality varies, and the report does not rank tools.
Is it illegal to make an AI copy of someone's singing voice?
It depends on where and how. Tennessee law covers voice simulations, and the Copyright Office recommends that federal liability arise from distributing or making available an unauthorized replica, not from creating one alone. No federal statute was in force at the last check.
What is a mel-spectrogram?
A representation of sound that shows how energy is spread across frequencies over time. DiffSinger produces a mel-spectrogram as its output.
Does the NO FAKES Act apply today?
No. It cleared the Senate Judiciary Committee on June 18, 2026, but had not passed either chamber, based on reports checked on October 27, 2026.
Related reading
- Whether AI can be creative
- AI-generated art
- AI paintings and how they are made
- Vectors in AI systems
More in AI Basics.
Last verified: October 2026
Sources
- Copyright and Artificial Intelligence, Part 1: Digital Replicas, U.S. Copyright Office, July 31, 2024
- Copyright and Artificial Intelligence landing page, U.S. Copyright Office, 2026
- Liu J et al., DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism, arXiv:2105.02446, accepted at AAAI 2022
- Gov. Lee signs ELVIS Act into law, Office of the Governor of Tennessee, March 21, 2024
- HB 2091 / SB 2096 (Public Chapter 588), 113th General Assembly, Tennessee General Assembly, 2024
- Rep. Salazar's NO FAKES Act advances out of Senate Judiciary Committee, Office of Rep. Maria Elvira Salazar, June 22, 2026