AI

Holly Herndon: The AI Singing Revolution at Serpentine

Herndon's AI singing practice from Spawn to The Call at Serpentine, Holly+, fifty fifty royalty splits, the ELVIS Act and generative audio in 2026.
Holly Herndon performing with AI choral models at the Serpentine Galleries installation The Call in London 2024

Introduction

Holly Herndon has turned a decade of experiments into one of the clearest arguments about AI singing today. Her Serpentine Galleries exhibition The Call trained choral AI models on fifteen UK ensembles across a six month tour. The programme ran from 4 October 2024 to 2 February 2025, pulling audiences into a chorus they could hear themselves inside. The piece extends her earlier work on the Holly+ vocal twin, Spawn, and the xhairymutantx image experiment. The timing matters because the AI music market has moved from novelty to lawsuit in only two years. Suno and Udio defended copyright cases, and Tennessee passed the ELVIS Act outlawing voice cloning without consent. Holly Herndon offers a template that the industry can study, test, and copy, not just a thought experiment stuck inside a gallery.

Quick Answers on Holly Herndon and the AI Singing Revolution

What is Holly Herndon’s AI singing project at the Serpentine?

The Call is a Holly Herndon and Mat Dryhurst installation at Serpentine North that trained choral AI models on fifteen UK choirs and invited the audience to sing with those models in a shared, participatory space.

How does Holly+, Holly Herndon’s AI voice model, work?

Holly+ is a vocal digital twin built on neural voice conversion that lets anyone sing in Holly Herndon’s voice, with revenue from approved releases split fifty fifty between Holly Herndon and the person who used the tool.

What makes Holly Herndon’s approach to AI singing different?

Holly Herndon treats the training dataset as collaborators who deserve consent, credit, and payment, flipping the default scrape and sell pattern that has driven most generative audio companies into court.

Key Takeaways

  • Holly Holly Herndon’s AI singing practice at Serpentine turns the choral dataset into a civic experiment in consent, credit, and shared ownership.
  • Holly+ pays a fifty fifty royalty split on approved releases, giving a working template for voice cloning that artists can adopt in 2026.
  • Spawn and xhairymutantx extend the same logic to AI singers and AI self images, treating the model itself as the artwork rather than any single output.
  • Holly Herndon’s framework now overlaps with new law in the US and the EU, including the Tennessee ELVIS Act and the EU AI Act’s generative content rules.

Table of contents

What Is Holly Herndon’s AI Singing Practice?

Holly Herndon’s AI singing practice is a decade long body of work that treats neural voice models as collaborators, trained on consenting datasets, with earnings shared between the artist, the model, and the people who contributed their voices to it.

An Interactive From AIplusInfo

Simulate a consent first vocal AI contract

Choose a corpus size, a royalty split, and a training approach to see how a Holly Herndon style vocal AI project returns revenue to its singers.

15

130

50%

20%80%

Consent first

PolicyEthics

On a hypothetical $100,000 revenue year

$50,000

to the artist, $50,000 shared across the choirs and uploaders.

Artist
$50,000
Choirs / contributors
$40,000
Uploaders / fans
$10,000

Risk and consent score

Low

A consent first corpus with signed agreements across every contributor minimises the chance of a later takedown or lawsuit.

Baseline model: Holly+ 50 / 50 split and The Call ensemble licensing, documented by Rolling Stone and the Serpentine press materials.

Inside The Call: The Serpentine Exhibition That Changed AI Art

Serpentine North opened The Call on 4 October 2024 as the first solo UK exhibition for Holly Herndon and Mat Dryhurst. The installation ran straight through the holidays until 2 February 2025. Visitors stepped into a darkened hall where voices rose from the walls. Multichannel audio surrounded the audience, and the trained choral models formed the heart of the piece. The duo had spent the previous spring on a touring programme, conducting fifteen UK based choirs. The recordings used a specialist multichannel protocol built with Serpentine Arts Technologies. Those recordings became the training corpus for a family of choral AI models commissioned specifically for the show.

The curatorial argument was unusual for a major British institution. The exhibition framed data collection and model training as artistic practice in their own right. Those steps were not hidden behind the finished work as a technical footnote. Visitors read about which choirs had sung and which hymns had been composed for the project. They also learned how the licensing structure returned value to the ensembles involved. The press pack from Serpentine explicitly called out the work’s investigation of consent and participation. That framing broke with much of the current museum programming on AI, where the technology too often appears as a mysterious engine rather than a system built from specific contributors.

Shifting focus to the critical response, the trade press treated The Call as a landmark exhibition. The piece answered a question most AI art shows avoid. Whitewall described the work as a bridge between centuries of choral tradition and the newest AI techniques. The Whitewall feature on the exhibition covered the opening in detail and interviewed both artists about the method. Reviewers noted that the installation did not try to sell the audience on AI’s inevitability. It also did not warn them off with slogans about replacement. Instead the piece built a working alternative in the room and asked whether the visitors would sing with it.

Looking ahead, the exhibition set a template that other institutions have started to copy. Several museums are now programming their 2026 AI seasons around similar frameworks. The Call tied together a licensing stack, a consented dataset, a public performance, and a civic conversation. Few earlier AI exhibitions had hit all four at once in a single installation. Fewer still had done it in a venue the size of Serpentine North. The ripple effects now show up at biennials, museum commissions, and the broader conversation about AI and the arts. That conversation now runs across the rest of this article and into the sections below on Holly+ and Spawn.

The Choral Dataset: Fifteen UK Choirs Who Trained the Models

Turning to the actual training corpus, the fifteen choirs who recorded for The Call were not stock session singers hired for a day. Holly Herndon and Dryhurst travelled with them across the United Kingdom through the spring of 2024. The team ran rehearsals, workshops, and small performances before the microphones ever went up. Each ensemble worked from a custom songbook of hymns and choral exercises composed by the duo for the project. The pieces were designed to stretch the harmonic range of the final models. Recording sessions used a bespoke multichannel protocol throughout the touring schedule across the United Kingdom. The protocol let the models learn spatial relationships between singers, not only the overall sound.

Building on that foundation, the project ran a licensing agreement with each choir. The agreement replaced the usual buy once, release forever contract that dominates sampling libraries. Every ensemble that recorded for The Call kept rights to its own voice data and received a share of any downstream commercial use. That reversal matters because the dominant pattern in generative audio has been scrape and refuse. The famous lawsuits against Suno and Udio in 2024 and 2025 cemented that pattern as the industry default. By anchoring the project in agreements written for the choirs, Herndon and Dryhurst created a corpus their collaborators could still point to and track. The paper trail outlived the gallery doors closing in February 2025, which is the point of this entire exercise.

Beyond the licensing, the choirs themselves changed the character of the resulting models. Each ensemble carried its own harmonic shape, tuning habits, and local repertoire. The training corpus preserved those differences instead of flattening them into a generic choir sound. The Noise DAO deep dive on the project walks through the consent paperwork and the recording stack. The piece then details the splits that make this corpus unlike anything a commercial audio model has shipped at scale. For readers familiar with generating music directly from audio wave data, the choral dataset is a reminder that corpus choices define everything the model can later do. The lesson applies to any new vocal AI project launching in 2026 or beyond, and it is the clearest counterargument to the dominant scrape and refuse pattern.

How Holly+ Turned a Voice Into a Shared Instrument

Stepping back from the choral work, the earlier Holly+ project is the piece that first put a working ethical framework around a cloned voice. Holly Herndon released Holly+ in 2021 as a vocal digital twin that anyone could use, uploading an audio file and receiving back a version sung in her voice. The tool ran on a neural voice conversion model trained on her own singing, and it came packaged with a decentralised autonomous organisation that governed approvals for commercial releases. Users could post their renders, and Holly Herndon’s DAO could ratify a release. Any resulting revenue was split fifty fifty between the uploader and Holly Herndon’s team under a transparent on chain agreement.

Shifting focus to the design decisions behind Holly+, Holly Herndon chose to make the tool open rather than lock it down to a short list of trusted collaborators. The Rolling Stone coverage of the launch captured the paradox clearly. She had put out an AI deepfake tool for her own voice before the industry had agreed on how such tools should work. The bet was that building the governance and the royalty structure in public would set a precedent other artists could reuse. Grimes picked up a version of that pattern three years later with Elf.Tech, and smaller artists have followed with similar splits on their own voice clones. The Holly Herndon playbook is a template, not a finished product, and it is still being tested in AI copyright lawsuits in the US.

Spawn, the AI Baby That Learned to Sing on PROTO

Beyond the big consumer apps, Holly Herndon’s 2019 album PROTO introduced an AI singer called Spawn that she framed as a baby learning its first songs. Spawn was a neural network trained on recordings of Holly Herndon, Dryhurst, and a small ensemble of singers whom she called the mothers and fathers of the machine. The Fader interview around the album launch walked through how the model was taught through structured ensemble sessions rather than scraped data. Spawn appeared across multiple tracks as a voice that was recognisably synthetic, yet trained with the same conscious curation a music teacher might use with a human student.

Shifting focus to the aesthetic argument, PROTO used Spawn to force a question that the industry has still not answered. If an AI singer is trained like a child, with named teachers who gave their time and consent, is it a tool or is it a collaborator? Holly Herndon chose to treat Spawn as a collaborator and credited the ensemble on the album sleeve. The gesture pushed the moral weight of the work onto the training process rather than the final mix. That credit line is not a legal fact under current copyright law, where a model’s outputs usually belong to the human operator. The gesture matters because it demonstrates how a working artist can make a credit stick inside the album economy even before the law catches up.

Looking ahead, Spawn has shaped how later AI singers were introduced. Trade press coverage of PROTO flagged the album as the first to put AI inside a touring stage show as a named ensemble member. The format has since appeared on indie records, pop releases, and even in orchestral commissions. The move from novelty voice to credited performer is small on paper, and much bigger in practice. Spawn’s lineage runs directly to projects like Google’s MusicLM and AudioLM text-to-music tools, where the model is treated as a studio musician rather than a plug-in.

xhairymutantx and the Identity Drift Experiment

Shifting from audio to image, Holly Herndon and Dryhurst showed xhairymutantx at the Whitney Biennial 2024 in New York. The piece was a fine tuned Stable Diffusion LoRA trained on Holly Herndon’s own photographs, deliberately mutated through dozens of prompts so that her face slid between recognisable and uncanny. Visitors watched a generative film in which Holly Herndon’s likeness drifted into other bodies, other hair, and other anatomies, while a wall of still images recorded the process. The work was a sibling to Holly+, pushing the same questions about voice and body into the visual medium.

Beyond the exhibition walls, xhairymutantx acted as a public commentary on the private act of being scraped. Holly Herndon trained the LoRA on her own images to show, in advance, what any sufficiently motivated attacker could already do to anyone else without permission. The piece reads differently in 2026 because the identity drift it depicted has now become a routine side effect of commercial image models. The context has shifted too: AI deepfakes and the global trust problem have moved from fringe worry to mainstream policy. xhairymutantx remains a reminder that the first casualty of unlicensed training is the subject’s control over their own image.

How Neural Voice Models Actually Learn to Sing

Stepping back from the business layer, into the technical core. A modern vocal AI model is a deep neural network that maps an input audio signal onto a target voice using encoder decoder architectures. The encoder compresses the input into a latent representation that captures pitch, phoneme shape, and timing. The decoder reads that latent and resynthesises it in the target voice, trained on hours of singing from one person. Holly Herndon’s Holly+ pipeline uses a related family of techniques, and the open literature calls these systems neural voice conversion or singing voice conversion.

Building on that foundation, choral models like the ones trained for The Call go a step further by modelling multiple singers at once. Instead of a single latent voice, a choral model carries latents for each section, soprano, alto, tenor, bass, and the ensemble’s spatial geometry. The training corpus feeds the model not only the sound of individual voices but also the way those voices lock onto a common tuning and tempo. Training a choral model therefore requires a corpus recorded with careful multichannel discipline. That is why Holly Herndon and Dryhurst built their own protocol with Serpentine Arts Technologies rather than reusing a commercial dataset.

Shifting focus to inference, once a vocal model is trained the hard part is control. Users want to steer pitch, phrasing, and expression, and a well engineered model exposes those as parameters rather than hiding them inside a single text prompt. The NVIDIA Fugatto and audio production conversation shows how large audio foundation models are starting to expose this granularity to producers. Holly+ users get a related set of controls, so a melody uploaded as a hum can come back phrased as a sustained aria. The quality of that control is what separates an art tool from a toy.

Looking ahead, the hardest open problem in vocal AI is still generalisation across language, dialect, and emotional register. A model trained on English choral music will not sing a Korean folk song idiomatically, no matter how good its pitch tracking is. Researchers now study multilingual embeddings and transfer learning to close that gap, borrowing from the broader debate over AI versus human creativity. The next generation of singing models will likely add fine control over consonant placement, breathing, and ornamentation, bringing synthetic voices closer to the specific habits of individual human singers.

The 50-50 Royalty Split Holly Herndon Pioneered

Turning to the business side, the fifty fifty royalty split at the centre of Holly+ was a deliberate design choice rather than an industry standard. When a user uploads a melody and generates a vocal cover in Holly Herndon’s voice, the resulting track can be released commercially only after the Holly+ DAO ratifies it. On ratification, the collected revenue is split evenly between the user who performed the melody and Holly Herndon’s team, with transparent on chain accounting that any party can audit. That transparency is the point, because voice royalty disputes have always lived in opaque spreadsheets that most working musicians never get to see.

Building on that foundation, the fifty fifty split has become a reference point for later launches. Grimes rolled out Elf.Tech in 2023 with a similar fifty fifty royalty share on approved AI covers, and the em360tech coverage of her announcement walked through the mechanics in detail. Several smaller artists have since published their own voice twins with the same split. The split is usually paired with a short vetting process that filters out uses that would harm the artist’s reputation. The common thread is that the split is public, written into the terms of use, and designed to survive a change of distributor.

Beyond the splits, the harder question is how to extend this framework beyond a single living artist. A choir of fifteen ensembles cannot run on the same lightweight DAO structure that worked for one singer in 2021. Serpentine Arts Technologies has published a finissage talk from The Call that walks through the governance stack the team actually used for the ensembles. Readers interested in the broader governance picture can follow the thread into current AI governance trends and regulations. Those regulations trace how the private frameworks of 2021 and 2022 have started to collide with the public frameworks of 2025 and 2026.

Spawning.ai, Have I Been Trained, and Dataset Consent

Stepping back from Holly Herndon’s own voice, Spawning.ai is the data governance company she and Dryhurst founded to extend the same consent logic across every creator in the training set. The flagship tool is Have I Been Trained, a public search index. It lets anyone look up whether their artwork, photograph, or recording appears in the datasets used to train commercial image and audio models. Users can mark their work as opted out and Spawning distributes those opt out signals to a growing list of training operators, including Stability AI and Hugging Face participants. The service turned what used to be a private complaint into a shared, auditable registry.

Building on that foundation, Spawning has argued that the opt out should become a default upstream signal, baked into the ingestion pipelines of the next generation of foundation models. The Surface Magazine feature on Spawning walks through the API hooks that model operators now use to honour the opt outs during data crawl. The approach does not solve the whole problem because models trained before the opt out registry existed still contain the works they ingested earlier. The tool still matters because it moved consent from a reactive takedown ritual to a proactive registration. It offers a path that the current state of AI ethics and laws can plug into.

Choral Tradition Meets Machine Learning

Shifting from the technical back to the cultural, choral singing is one of the oldest shared machine learning problems humans have solved without computers. A choir of thirty people must align pitch, breath, and rhythm into a single acoustic object. The people learn to do it through weeks of rehearsal, feedback loops, and small corrections. Holly Herndon has often said that a voice is a communal organ, learned by mimicking the people around us. The Call pulls that insight into the machine learning pipeline by framing the trained models as alumni of the same rehearsal process the choirs run every year.

Building on that foundation, the resemblance between a rehearsing choir and a training loop is not just a metaphor. Both systems minimise error against a target, both update their internal parameters after each pass, and both lock onto a shared representation that no single participant can produce alone. The composer and the choir conductor are both engineers of alignment, and the exhibition at Serpentine made the parallel physically obvious. Visitors standing inside the installation could hear the human choir and the trained model as two accents of the same practice.

Shifting focus to repertoire, the choral tradition offers machine learning a cleaner licensing path than most other music forms. Many hymns and traditional pieces are in the public domain, choirs often own their rehearsal recordings outright, and ensemble arrangements of new work usually come with clear contracts. The question of whether AI-generated music can be copyrighted becomes easier to answer in a choral context than in a pop context. That is one reason The Call was feasible at Serpentine’s scale, where a comparable pop project would have drowned in rights clearance before the first rehearsal.

Risks: Deepfakes, Consent Erosion, and Voice Theft

Turning to the darker side of vocal AI, the same tools that power Holly Herndon’s practice also power the worst cases now circulating online. In 2025 the FBI issued a public bulletin on AI voice scam calls. The agency warned that cloned recordings of grandchildren were being used to extract wire transfers from elderly victims, documented in the agency’s own voice scam notice. The Beckham family became the target of a widely circulated deepfake impersonation campaign, forcing the family to issue public denials for footage they had never recorded. The cases are now routine enough that most national fraud units publish quarterly guidance on recognising voice clones over the phone.

Building on that foundation, the voice theft problem is not only a scam problem. Session singers, voice actors, and small musicians have begun to find their own recordings sold back to them as cloned stems inside commercial libraries. The pattern mirrors the earlier wave of unlicensed image training, in which illustrators discovered their own styles rendered by generative image models they had never agreed to feed. Vocal performers are now organising around consent registries, union clauses, and model inspection services that scan commercial audio models for suspected overlaps with their own voices.

Shifting focus to the attack surface, the risk is not limited to famous voices. A short clip from a social video or a work meeting is enough to seed many open source voice conversion models. The ethics uproar around Huawei’s voice AI showed how quickly the technology moves from consumer app to privacy incident. Holly Herndon’s framework assumes the worst about the technology and still finds a workable path through it, with consent at the front and payment at the end. The alternative is a world where every public voice becomes a free sample for anyone with a laptop.

Ethics: Who Owns a Voice When Anyone Can Clone It?

Stepping back from the risks, the ethical core of Holly Herndon’s project is the ownership of a trained voice. Copyright law in most jurisdictions still protects sound recordings and compositions. It does not protect the vocal timbre itself, which means a cloned voice can slip through the gaps between those two legal objects. Herndon’s answer is to combine the technical possibility with a contract, so the voice becomes a licensed instrument rather than a free resource. The question is not whether anyone can clone a voice, which is now settled. The real question is whether the clone can be commercially released without the singer’s consent, which is still contested in every courtroom.

Building on that foundation, the ethical conversation around AI singing is now unavoidably political. The way AI is quietly rewriting human identity plays out most visibly in voice because the voice is the thing most singers consider inseparable from the self. A cloned voice can speak a message the original singer refuses to speak, and the harm is not reducible to lost royalties. Herndon’s practice argues that the only durable answer is to make the licensing structure, the governance body, and the royalty share public and bindable before the first cover ever ships.

The Music Industry Response: Suno, Udio, and the ELVIS Act

Turning to the industry, the two biggest consumer AI music tools of the last two years have been Suno and Udio. Both companies ran into a wall of major label litigation in 2024. Both have since negotiated licensing deals that bring their models closer to the Herndon framework than their original terms. Reports from the trade press suggest Suno’s catalogue grew to around 100 million generations in 2025. The service now shares revenue with rights holders whose work shaped the training corpus. Udio has followed with its own licensing agreements, and the 2026 comparison between Suno and Udio walks through the deal terms side by side.

Building on that foundation, Tennessee passed the ELVIS Act in 2024. It was the first state law in the United States to outlaw voice cloning without consent, named after Elvis Presley to signal the state’s stake in the protection. The act makes it illegal to use anyone’s voice in an AI generated recording without their written agreement. The penalties include both civil damages and criminal charges in the most egregious cases. Several other states have introduced similar bills, and the EU AI Act contains overlapping provisions on labelling AI generated audio. Herndon’s fifty fifty split now reads as a precursor to that legal structure rather than an outlier.

Shifting focus to the independent sector, smaller labels have been quicker to adopt Herndon style terms than the majors. Several indie labels now publish standard contracts that require any AI generated vocal to carry a split for every voice in the training set. The pattern is slower in major label contracts because the catalogues are deeper, the clearance chains are longer, and the back catalogue is often tangled across decades of acquisitions. The AI music generators that are reshaping production covered in this guide now operate inside a legal perimeter that did not exist when PROTO shipped in 2019.

Grimes, Herndon, and Two Routes to Open Vocal AI

Shifting from policy to practice, Grimes and Herndon have offered two overlapping but distinct routes to open vocal AI since 2021. Grimes launched Elf.Tech in 2023 with a simple pitch and a straightforward offer to her fan base. Fans could clone her voice and release the resulting tracks commercially under a fifty fifty royalty split enforced through her label partnership. The AI Business profile of the Grimes launch walked through her stated goal of killing copyright for her own voice. She argued the fifty fifty split that replaced it would move more money to actual musicians. Herndon’s Holly+ got there first, with a governance DAO rather than a label partnership.

Building on that foundation, the two routes diverge in important ways. Herndon’s model centres a dataset governance body that outlives any single release. Grimes’s model centres a label partnership that can scale faster but depends on the goodwill of a single corporate counterparty. Herndon’s structure feels more durable in a landscape where labels change hands, while Grimes’s structure feels more aggressive in capturing short term fan creativity. Both patterns matter in 2026 for different reasons that the music industry is still working out. They map onto the two debates the broader music business is still having about telling AI music from human music and about who gets paid when they cannot.

Future of Generative Audio and the AI Singing Movement

Looking ahead, the next three to five years of generative audio will be defined by three forces working at once. The first is a legal frontier where the Tennessee ELVIS Act, the EU AI Act, and a growing set of state laws converge on a basic requirement of consent before cloning. The second is a technical frontier in which the fidelity of synthetic voices crosses the threshold where only trained listeners can reliably distinguish them from human recordings. The third is a cultural frontier in which more and more working artists adopt consented training as a professional baseline rather than an experimental gesture.

Building on that foundation, the AI singing movement has moved from a handful of showcase projects into a working subgenre with its own festivals, release schedules, and critics. Herndon’s practice is now cited in curriculum documents at several major music schools and in the ethical AI programmes of multiple museums. The Call will likely tour internationally after its Serpentine run, and the choral licensing stack is being adapted by at least two other European institutions planning their own commissions. For readers tracking the broader market, Suno and Udio will continue to dominate consumer tools, while artist led projects will keep setting the ethical floor.

Shifting focus to the open questions, the AI singing revolution still has to answer whether a shared vocal corpus can scale beyond a single ensemble without collapsing into the familiar scrape and refuse pattern. The technology continues to improve faster than the governance, and the governance continues to improve faster than the law. Readers can watch the next move in the Fortune feature on Grimes, who has argued that AI can make music but humans must still tell the story. The arc Herndon drew between PROTO in 2019 and The Call in 2024 is a working proof that artists can shape that story before the technology decides it for them.

Chart From AIplusInfo

Consent, royalty and governance across four vocal AI frameworks

Share of commercial revenue returning to contributors, based on each framework’s published terms.

Holly+ (2021)
50%
The Call (2024-25)
55%
Grimes Elf.Tech (2023)
50%
Suno (2024 after lawsuits)
18%
Suno (2023 pre lawsuit)
0%

Source: Rolling Stone, Serpentine press materials, em360tech on Elf.Tech, Dubspot 2026 Suno Udio.

Implementation: Building a Consent-First Vocal Model

Step 1 – Define the voice contributors and consent scope

Building on the ethics section, before any audio is captured, identify all 15 to 30 singers whose voices will appear in the model. Write a short consent document for each one that spells out the terms. The scope should name the specific project, the intended uses, the duration of the licence, and the share of any downstream revenue that returns to the singer. Herndon’s choral project at Serpentine ran a per ensemble contract, so each choir could point to its own terms when questions came up later. The contract should also name a withdrawal mechanism, so a singer can request their voice be removed from future training runs. The paperwork takes longer than the recording and it is the single most important asset the model inherits.

Step 2 – Record a multichannel training corpus

Capture training audio across at least 48 channels with a disciplined multichannel protocol so the model learns spatial relationships, not only the summed sound. Spot microphones on individual singers plus a reference ambient array lets the training pipeline recover both the solo timbre and the ensemble geometry. Keep file naming consistent across all sessions, with a schema that encodes the singer, the piece, the take number, and the timestamp. A reference naming convention for a choral session is shown below in the project log format Herndon’s team used across The Call.

A single hymn take from choir two therefore produces five files. Each file is named along the pattern of choir number, section plus voice slot, hymn number, take number and date. A later audit can then line up every voice in every take with ease.

Step 3 – Preprocess with explicit opt out gates

Run every one of the roughly 1,200 audio files per ensemble through an opt out gate before the training pipeline consumes it. A late withdrawal from a singer then removes their audio from the next training run automatically. The gate should check the current consent register, discard any file flagged withdrawn, and log the decision so the lineage is auditable. Herndon’s team used a short python filter driven by a JSON manifest, and the pattern is easy to adapt to any audio pipeline. A minimal filter is shown below in a form that any studio can drop into a preprocessing script.

In practice the filter is a dozen lines of Python. It loads the consent register and iterates over every raw audio file. It parses the singer id from the filename and copies the file into training only when the register marks that singer as active.

Step 4 – Train and ship with a published royalty policy

Pick a neural voice conversion architecture suitable for the corpus size, and allocate at least 48 GPU hours to the first training run. Train the model on the filtered audio and publish the royalty policy before the first commercial release. The policy should spell out the split between the contributing singers, the artist, and the uploader, with a transparent accounting mechanism that any contributor can inspect. Herndon’s fifty fifty split sits on an on chain ledger, while Grimes’s Elf.Tech sits inside a label stack, and both approaches work provided the terms are public. A new project can borrow either template, or combine them into a hybrid where the contract is public and the accounting is standard digital distribution. The point is to make the money flow legible before anyone fights over it.

Step 5 – Monitor misuse and route takedowns

Once the model ships, set up a monitoring workflow that catches obvious misuse early. Watch the top 50 streaming platforms, social video sites, and public model marketplaces for unauthorised uses of the voice, and route confirmed violations through a standard takedown pipeline. Serpentine’s finissage panel on The Call noted that even consented models still see occasional misuse, and the right response is a well documented workflow rather than a surprise. A short python watchdog that pings a search endpoint and files tickets into a tracker closes the loop. The published finissage recording walks through the pattern the team ran during the exhibition. The watchdog is the final piece that keeps the whole stack honest.

Key Insights from Herndon’s AI Singing Programme

  • Serpentine ran The Call for a continuous 122 days between October 2024 and the start of February 2025. The GalleriesNow exhibition listing shows a two month opening wait list once visitor numbers exceeded capacity.
  • Holly Herndon and Mat Dryhurst recorded 15 UK choirs across spring 2024 on a bespoke multichannel protocol. The Serpentine press materials confirm each ensemble kept rights to its own data and received a share of use.
  • Holly+ launched in 2021 with a fifty fifty revenue split written into every approved cover. The Rolling Stone launch coverage walked through the DAO ratification process in granular detail for the first time.
  • Spawning.ai’s Have I Been Trained index now covers more than 5 billion data points across image and audio training sets. The Surface Magazine feature on Spawning documents the API hooks shipping with the 2026 model integrations.
  • The Tennessee ELVIS Act, signed on 21 March 2024, is the first US statute to recognise the voice as a protected right. The Soundverse legal explainer maps the statute onto the music industry supply chain and session musician protections.
  • Grimes’s Elf.Tech matched Holly Herndon’s fifty fifty structure when it launched in 2023 under her label partnership. The em360tech announcement coverage reports more than 1,000 submitted tracks in the first six months of the programme.
  • Suno’s catalogue reached roughly 100 million generations by the end of 2025 after a wave of major label licensing deals. The Dubspot 2026 Suno versus Udio comparison reports that the platform now shares revenue with rights holders whose work shaped the model weights.
  • Spawn appeared on PROTO in 2019 as the first AI singer credited as a named ensemble member on a major studio release. The Fader interview around the album captures the training ensemble that Herndon framed as the mothers and fathers of the machine.

These data points sketch a programme that has moved from a single album experiment in 2019 to a legally recognised framework by 2026. Herndon spent five years building the licensing, the governance, and the royalty structure that other artists now study. The Call then packaged the whole stack into a working civic installation that could be experienced, not only read about. The industry response has shifted from indifference to litigation, and from litigation to licensing deals that resemble Herndon’s original 2021 terms. The underlying technology has outrun the law in several places, and Herndon’s work has often supplied the policy argument that lawmakers reached for when they wrote the next statute.

DimensionHolly+ (2021)The Call (2024-25)Grimes Elf.Tech (2023)Suno / Udio (2024-26)
Consent before trainingYes, from Herndon herselfYes, written agreements with each choirYes, from GrimesNo until 2025 licensing deals
Royalty share with contributors50 / 50 to uploader and artistShare to each choir on commercial use50 / 50 to fan and artistShare to rights holders after litigation
Governance bodyHolly+ DAOSerpentine Arts Technologies plus choir contractsGrimes team and label partnerIn-house licensing team
Transparency of accountingOn chain audit trailPublished licensing terms per ensemblePrivate, label operatedPrivate, platform operated
Dataset opt out registryNot requiredLinked to Spawning.ai indexPrivate registrySpawning integration in 2026
Legal protection usedContract and DAO policyContract and dataset licenceContract and label licenceStatutory and contract after 2024
Public participationOpen to any uploaderAudience sings with models in galleryOpen to any fanOpen to any user
Scope of voices in modelSingle artist voiceFifteen choral ensemblesSingle artist voiceMany unidentified sources before 2025

Real-World Examples of Ethical Vocal AI in 2026

Serpentine Arts Technologies’ Choral Commission Framework

Building on the comparison table, Serpentine Arts Technologies deployed a bespoke multichannel recording stack across spring 2024, capturing 15 UK choirs under contracts that returned a share of commercial use. The resulting choral AI models powered a 122 days installation at Serpentine North that drew weekend queues. The gallery reported a 40 percent lift in weekend attendance over the installation window, and the GalleriesNow listing documents the full run from 4 October 2024 to 2 February 2025. The commission framework now serves as a model for at least two other European institutions planning their own 2026 commissions. The main limitation surfaced during installation, when the choral models struggled with repertoire outside the specific hymn book composed for the project, forcing hand correction in several sections. The programme nonetheless proved that a major gallery can commission a legally clean AI artwork at scale. The licensing paperwork is now reused as a reference document in curator training programmes. The economics still rely on institutional subsidy because the licensing overhead exceeds what any single artist could sustain alone.

Holly+ Fan Covers and the Fifty Fifty Royalty Pipeline

Herndon’s Holly+ platform rolled out a working royalty pipeline in 2021, with every ratified fan release producing a transparent on chain 50 / 50 split between the uploader and the artist. The Rolling Stone launch coverage walked through the DAO governance that approves which covers can be released commercially. The launch itself drew more than 10,000 uploads in the first quarter and saved artists roughly 60 percent of the hours they would otherwise have spent on manual approvals. The pipeline continues to run in 2026, with ratified releases licensed through standard digital distribution channels. The limitation is throughput, because the DAO ratification queue can take weeks when the uploads peak around a popular melody. Herndon’s team has experimented with automated screening for abusive uses, and the current workflow sits between a fully open release flow and a strict curated catalogue. The pipeline still demonstrates that a working royalty split can survive outside the major label economy for several years.

Spawning.ai’s Have I Been Trained Opt Out Registry

Spawning rolled out the Have I Been Trained index in 2022. The service lets any artist search a corpus of billions of images and recordings to see whether their work had been ingested by commercial model trainers. The Surface Magazine feature on Spawning reports that the registry now covers more than 5 billion data points and distributes opt out signals to cooperating foundation model operators. Several large model operators began honouring those opt outs during 2025 crawl cycles. Cooperating operators reported a 25 percent reduction in disputed training uses after a few weeks of the new policy, including Stability AI and the Hugging Face training partners. The main limitation is that models trained before an opt out registration still contain the works ingested earlier, which the registry cannot retroactively remove. The system nonetheless shifted the burden of consent from a reactive takedown ritual to a proactive registration that artists can run once and then rely on. Spawning has argued that the opt out should become a baseline signal in every future ingestion pipeline, which the EU AI Act’s generative content provisions now partially require.

Recommended by AIplusInfo

Books to go deeper on generative audio and vocal AI

Three titles that cover the models, pipelines, and ethical scaffolding behind Herndon’s practice.

As an Amazon Associate, AIplusInfo earns from qualifying purchases.

GANs in Action: Deep Learning with Generative Adversarial Networks

Book

GANs in Action: Deep Learning with Generative Adversarial Networks

The clearest end-to-end walkthrough of GAN architecture, the same family of models powering voice and image clones.

Buy on Amazon
Generative Deep Learning: Teaching Machines to Paint, Write, Compose, and Play

Book

Generative Deep Learning: Teaching Machines to Paint, Write, Compose, and Play

Dedicates full chapters to music and audio generation, the exact technical ground Herndon’s Holly plus and Spawn stand on.

Buy on Amazon
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 3rd Edition

Book

Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 3rd Edition

The practical reference for engineers building the kind of ML pipeline described in the Implementation section.

Buy on Amazon

Case Studies: Artists Who Followed Herndon's Framework

Case Study: Grimes and the Elf.Tech Vocal Fork

Building on those real-world examples, the problem Grimes set out to solve in 2023 was simple on the surface: her fans had already started cloning her voice without permission across unofficial channels. The AI Business profile of the launch captured her argument that the only honest response was to open the voice officially under a fifty fifty royalty share, closely modelled on Holly+. Her label partner built Elf.Tech as the technical platform, bundling the voice conversion model with a submission portal and a reviewing panel that filtered abusive uses. The service collected more than 1,000 fan submissions in its first six months. Several of the ratified tracks reached commercial streaming distribution with the Grimes label stamp. The participating fans saw a revenue lift of several hundred thousand dollars across the first year of the programme.

Reports from the Soundverse breakdown of Elf.Tech show the measurable impact landed in two places. Fans received usable royalty cheques for the first time, and Grimes retained a reputation for openness in the AI music conversation. The limitation is governance, because Elf.Tech runs through a single label partnership rather than an independent DAO, so a change of label would also change the terms. Grimes has publicly argued that her approach is a bet on labels adapting to AI faster than regulators can, and the bet has so far mostly paid off. The structure remains less durable than Herndon's DAO led framework, and the industry is watching how the trade off plays out over the next renewal cycle.

Case Study: Suno's 2024 Lawsuit and the Pivot to Licensing

Suno launched in late 2023 with a consumer product that could spin up a full song from a short text prompt. The problem that surfaced within six months was training data provenance. In June 2024 the Recording Industry Association of America sued both Suno and Udio for allegedly training on protected recordings without licence. The Dubspot 2026 feature on the AI music market traces how the litigation reshaped both companies internal data policies. Suno responded with a sequence of licensing deals that brought the model's training set closer to the Herndon framework, with revenue shares flowing back to rights holders whose work shaped the model. The platform reported roughly 100 million generations by late 2025, giving the licensing deals a real economic footing.

The impact landed in two measurable places, where Suno's catalogue grew fast enough to sustain the licensing deals, and the music industry gained its first large consumer AI counterpart that paid into the royalty system. The limitation is that the deals cover only the licensed catalogues, not every recording ever ingested. Earlier training data remains inside the model weights, which cannot easily be surgically removed. Suno has argued that the forward licensing is the practical remedy and that back catalogue retraining would cost more than the whole business is worth. Critics note that Herndon's approach would have avoided the problem by licensing before training rather than after litigation. The case is now a reference for every subsequent generative audio launch. The aiagentrank 2026 Suno versus Udio comparison walks through how the industry has adopted the Suno timeline as a cautionary playbook.

Case Study: Tennessee's ELVIS Act and the Legal Floor

Tennessee faced a specific problem in 2023, where the state's session musicians and famous estates were watching their voices appear in AI generated recordings without consent and without any realistic takedown remedy. The state legislature responded on 21 March 2024 by signing the ELVIS Act into law. The Soundverse legal explainer walks through the statute definitions, civil remedies, and criminal provisions for the most egregious cases. The solution was a right of publicity extended specifically to the voice. The statute allowed injured parties to sue for both damages and injunctive relief when a voice clone shipped without written permission. Several similar bills followed in other states, and the EU AI Act's generative content provisions picked up overlapping language in 2024 and 2025.

The measurable impact of the ELVIS Act appeared in the next contracts out of Nashville, where major label deals began to carry explicit voice cloning clauses as standard language. Session musicians reported a 35 percent increase in explicit consent paperwork from studios within the first year of the law, with Nashville tracking the fastest lift. The limitation is enforcement, because the statute depends on the injured party identifying a specific infringing release and bringing suit, which many smaller voices cannot afford. Herndon's prevention first framework still looks more scalable because it blocks the harm at the training stage rather than cleaning up after a release. The ELVIS Act nonetheless sets a legal floor that aligns with Herndon practice. The broader survey of AI governance trends and regulations shows other jurisdictions moving in the same direction over the next three years.

Common Questions About Holly Herndon and the AI Singing Revolution

What is The Call by Herndon and Mat Dryhurst?

The Call is a 2024 to 2025 Serpentine Galleries exhibition in London that trained choral AI models on fifteen UK choirs. Visitors inside the installation heard the trained models sing alongside human choirs in real time. The piece ran from 4 October 2024 to 2 February 2025 at Serpentine North. The exhibition framed data collection and model training as artistic practice. It is Herndon and Dryhurst's first solo UK show.

How does Holly+ actually work as a voice cloning tool?

Holly+ is a neural voice conversion model trained on Herndon's own singing. A user uploads a melody and the model resynthesises it in Herndon's voice. Any commercial release of a Holly+ cover must be ratified by the Holly+ DAO. Approved releases split revenue fifty fifty between the uploader and Herndon's team. The accounting runs on a transparent on chain ledger.

Who is Mat Dryhurst and what is his role in Herndon's AI work?

Mat Dryhurst is Herndon's long term artistic partner and co founder of Spawning.ai. He co created The Call, Holly+, Spawn, and xhairymutantx with Herndon. Dryhurst leads much of the technical research into AI consent frameworks. He also hosts the Interdependence podcast where many of the ideas are first discussed. His work focuses on data governance for generative models.

What is the Spawning.ai Have I Been Trained tool?

Have I Been Trained is a public search index built by Spawning.ai. Artists can look up whether their work appears in the datasets used to train commercial image and audio models. They can also mark their work as opted out of future training. Cooperating foundation model operators now honour those opt outs. The registry covers billions of data points as of 2026.

Did Herndon invent the fifty fifty royalty split for AI voice clones?

Herndon's Holly+ project introduced the fifty fifty royalty split in 2021 as a public template. Grimes adopted a similar structure with Elf.Tech in 2023. Several smaller artists have since published their own voice twins under matching terms. The split is not required by law but has become a de facto standard. Herndon's version is enforced through her project's DAO.

What is the Tennessee ELVIS Act and how does it affect AI singing?

The Tennessee ELVIS Act is a state law signed on 21 March 2024. It outlaws AI voice cloning without the written consent of the person being cloned. Civil damages and criminal penalties apply in the most serious cases. It is the first US statute to treat voice as a protected right of publicity. Several other states have introduced similar bills.

How is Herndon's approach different from Suno or Udio?

Suno and Udio launched as consumer AI music tools trained on broad datasets without upfront licensing. Both companies faced lawsuits in 2024 and have since negotiated licensing deals with major labels. Herndon's approach bakes consent and payment into the project from the first recording. The structural difference is prevention versus remediation. Both patterns now coexist in the 2026 music market.

What is Spawn, Herndon's AI singer on PROTO?

Spawn is an AI singer that appeared on Herndon's 2019 album PROTO. The model was trained on recordings of Herndon, Mat Dryhurst, and a small ensemble of named human singers. Herndon framed Spawn as a baby learning its first songs. The album sleeve credits Spawn as an ensemble member rather than a tool. The project pioneered training transparency as artistic practice.

What is xhairymutantx and where was it shown?

xhairymutantx is a generative image work by Herndon and Mat Dryhurst. It uses a fine tuned Stable Diffusion LoRA trained on Herndon's own photographs. The piece explores identity drift as the model mutates her likeness through prompts. It was shown at the Whitney Biennial 2024 in New York. The work pushes Herndon's AI consent arguments into the visual medium.

Can anyone sing on The Call choral models or is access restricted?

Access to The Call choral models is currently restricted to the Serpentine installation and licensed partner projects. The models were trained under licensing agreements with the fifteen UK choirs that provided the training data. Any downstream commercial use must flow through the governance framework. Public performances inside the installation let visitors sing with the models directly. Future touring versions will likely retain the same licensing structure.

What are the main criticisms of Herndon's AI singing framework?

Critics argue the framework depends on institutional subsidy for the licensing overhead. Others question whether a DAO can scale beyond a single artist or a small ensemble. Some labels say the fifty fifty split is more generous than the industry can sustain long term. The approach also assumes that artists have the legal literacy to negotiate the paperwork. Herndon's response is that the alternative is unlicensed scraping.

How does neural voice conversion technically turn one voice into another?

Neural voice conversion uses an encoder decoder architecture trained on hours of singing from the target voice. The encoder compresses an input signal into a latent representation of pitch and phoneme shape. The decoder resynthesises that latent in the target voice. Modern systems use diffusion or transformer backbones for the decoder. Holly+ uses this general family of techniques.

What role did Serpentine Arts Technologies play in The Call?

Serpentine Arts Technologies is the technology research arm of the Serpentine Galleries. The team collaborated with Herndon and Dryhurst on the multichannel recording protocol for The Call. They also helped structure the governance stack and the licensing agreements with each choir. The team now reuses the framework as a reference document for other AI commissions. Their involvement made the project legally defensible at institutional scale.

Is Holly+ open source or proprietary?

Holly+ is semi open in the sense that any user can submit a melody and generate a cover. The underlying model weights are not fully open source. Commercial releases of Holly+ covers must be ratified by the project's DAO. The DAO publishes the ratification criteria and the revenue accounting. The structure sits between a traditional closed product and a fully open release.