Voiceprint Lawsuits Put AI Training Permissions Under Scrutiny

Nine related class actions filed in May 2026 challenge the use of Illinois voice professionals’ recordings to develop commercial voice AI. The plaintiffs allege that companies extracted biometric voiceprints from publicly available audio without the notice, permission and safeguards required by the Illinois Biometric Information Privacy Act. The claims put a question about data transformation alongside the familiar debate over AI training rights: what, exactly, did a company extract from a recording?

In an October 20 IAPP analysis, Ashurst Perkins Coie Counsel Anna Onley, AIGP, and Chicago Nova Opera Legal Fellow Ganna Yelovets connect the litigation to governance throughout the life of voice data. Their introduction examines the records needed to establish how recordings were obtained, transformed and used, rather than treating an initial consent decision as the end of the review.

For businesses, that is a concrete documentation problem. A recording may enter through a licensed production, a customer call or a public dataset. Later processing can separate the spoken words from characteristics associated with the speaker. The permissions and controls supporting the original recording need to be examined before that new use proceeds.

Case and statutory materials reviewed October 23, 2026. Allegations, procedural rulings and practical recommendations are distinguished below.

The May lawsuits target biometric extraction from public audio

The coordinated filings include Marin v. Alphabet, Amer v. Eleven Labs, Rogers v. NVIDIA and actions against Amazon, Apple, Adobe, Meta, Microsoft and Samsung. An American Bar Association overview identifies the nine cases, filed May 11–14 in the Northern District of Illinois.

The Marin complaint names seven plaintiffs, including broadcast journalist Carol Marin and audiobook narrators Lindsey Dorcus and Victoria Nassif. It alleges that Google used their recorded voices in developing voice AI and that none received the required notice or executed a written release. It also alleges that voice-derived information became incorporated into model parameters. Those are plaintiffs’ assertions, not established findings about Google’s architecture or conduct.

The complaints reach beyond copyright theories. A license to distribute an audiobook and permission to extract a speaker-identifying representation concern different rights. The litigation asks whether the defendants obtained covered biometric data and, if so, met the applicable requirements. Public availability does not answer those questions by itself.

The defendants have contested the claims. Reuters reported in August that defense arguments included whether the alleged conduct occurred in Illinois. The location and nature of each relevant operation therefore matter alongside the characteristics of the data. No liability judgment or settlement resolving the May cases was verified for this article.

A voice recording and a voiceprint require separate analysis

BIPA expressly lists voiceprints as biometric identifiers. It separately defines biometric information by reference to an identifier used to identify an individual. The statute does not make every audio file or transcription a voiceprint.

Technical review should identify what the system creates and what it can do with it. A transcript represents words. A speaker embedding may encode characteristics useful for matching recordings to a speaker. A synthetic voice model may reproduce vocal attributes. Those descriptions do not settle the legal classification, but they identify the evidence counsel needs.

Document whether representations persist across sessions, whether the system compares them with other recordings and whether account information links them to a person. Removing a name from a filename does not resolve identifiability if the surrounding system retains a usable speaker mapping. Conversely, an allegation that software analyzes audio does not establish every element of a biometric claim.

The distinction appeared in Delgado v. Meta Platforms. An order publicly filed June 25, 2026 denied Meta’s motion for summary judgment. Meta argued that ordinary recordings were outside BIPA and that the recordings did not enable it to identify the plaintiff. The court found a material factual dispute concerning identification using technology Meta possessed.

The court expressly declined to define the precise point at which a recording becomes a voiceprint. Its decision was not a finding that Meta violated BIPA. For compliance teams, it is a reason to investigate actual technical capabilities rather than rely solely on a feature’s stated purpose or its “transcription” label.

Voice ID disclosures can become a class-wide question

An earlier Alexa case shows why the wording presented to users matters. In Gunderson v. Amazon.com, on the docket originally captioned Wilcosky, the court certified a class, with a public version of its opinion filed November 19, 2025. The dispute concerned Alexa Voice ID enrollment and alleged failures under BIPA.

Amazon maintained that its disclosures supported informed consent and challenged whether the feature created covered identifiers. The court treated common questions about the enrollment process and disclosures as suitable for class treatment. Certification did not decide the merits.

The operational implication is to preserve the actual notice, release and enrollment version. A current help page cannot establish what a user saw years earlier. Nor should a team assume that a description of personalization necessarily explains every biometric operation involved in delivering it.

BIPA imposes duties beyond obtaining a release

Section 15 addresses several distinct obligations. Section 15(b) requires prior written notice of collection or storage, the specific purpose and duration, and a written release. Section 15(a) requires a public retention and destruction policy, with destruction when the initial purpose is satisfied or within three years of the last interaction, whichever occurs first. Section 15(d) restricts disclosure, subject to exceptions; § 15(e) addresses protection.

Section 15(c) prohibits selling, leasing, trading or otherwise profiting from covered biometric data. Consent should not be treated as a universal answer to that separate provision.

The Act permits damages of $1,000 for negligent violations or $5,000 for intentional or reckless violations, or greater actual damages, plus potential fees and other relief. Its 2024 amendment limits specified repeated collections and disclosures to single recoveries under defined conditions. Exposure should not be calculated automatically per recording.

For a voice project, separate the authorization for recording from any required biometric release and from the rights needed for the intended AI use. A supplier’s permission to deliver a dataset does not establish that each speaker authorized every proposed transformation. The review should identify the source and scope of the authority being relied upon.

Build permissions into the path from recording to training

Start with the acquired material. Record the source, acquisition date, supplier and permitted uses, alongside evidence supporting those permissions. Preserve relevant restrictions on speaker identification, synthetic voice creation and onward sharing. A dataset described as suitable for research should not silently become approved material for a commercial voice product.

Track the transformations that follow. For an illustrative pipeline, a recording might be segmented, transcribed, assigned a speaker label and converted into an embedding before entering a training set. Connect those derived assets to their source records. Otherwise, deleting or restricting the original file may leave the useful speaker representation available elsewhere.

Approval should attach to a defined use. Improving transcription accuracy, authenticating callers and generating a recognizable synthetic voice present different questions. A change between those uses should trigger review before the relevant data moves into the new workflow.

Captain Compliance’s data mapping and AI Governance software support records connecting systems, purposes, vendors and responsible owners. Applied here, that structure can document which voice-processing use has been approved and the evidence behind it. Assessment workflows can route unresolved questions to the appropriate reviewers and track required mitigation.

The implementation still needs to enforce those decisions. A training job should not accept excluded material merely because an approval record exists elsewhere. Teams need a reliable mechanism for propagating restrictions into dataset selection, access permissions and vendor instructions, followed by checks that the restrictions took effect.

Vendor access and deletion need evidence of their own

A voice pipeline may involve a recording supplier, annotation service, model developer and hosting provider. Identify which party receives raw audio and which receives derived representations. Contracts should address approved uses, retention and onward disclosure in terms that match those transfers.

Technical records should make it possible to distinguish a file transfer from remote access or a model-training operation. Review whether development tools, shared workspaces or diagnostic logs create additional copies. A vendor’s general assurance that information is confidential leaves unanswered whether it may use the recordings to improve its own products.

Deletion becomes harder once data has produced derivatives. Inventory audio, speaker mappings, embeddings, training subsets, checkpoints and deployed versions as relevant to the architecture. Their legal treatment requires analysis; they should not all be assumed to be biometric data merely because audio appeared upstream.

The Marin allegations make that distinction particularly important: the plaintiffs contend their voice-derived information remains in model parameters. Whether that assertion is proved, and what remedy would follow, remain separate questions. An enterprise should not promise that deleting a source recording reverses an earlier training operation.

Where a deletion or restriction is required, establish what can be removed directly and what needs further technical and legal review. Preserve evidence of completed actions and any unresolved dependencies. If retraining or another model-level intervention is proposed, test its effect before describing the issue as resolved.

Research also supports examining later reuse rather than only initial collection. Tanusree Sharma, Yihao Zhou and Visar Berisha’s 2025 study, based on interviews with 20 professional voice actors, describes concerns about recordings repurposed for synthetic voices and the resulting privacy, reputational and accountability risks. Those interview findings are not evidence against the defendants, but they help explain why an old recording can raise a new governance question.

Test whether one voice can be accounted for

A useful review can begin with one authorized sample. Ask the team to locate its source permission, identify each derived representation and show which datasets and model versions received it. Then introduce a restriction: the sample must no longer be used for a particular purpose. Check which systems can enforce that decision and which cannot.

The exercise may expose a missing supplier record, a reusable speaker profile or a training export whose contents no longer have traceable permissions. Assign those findings before the next dataset refresh. That is where these lawsuits become operationally useful: they give teams specific claims to test against their own records.

The next release decision should name the voice data it depends on and the authority supporting that use. If either cannot be established, the affected material needs to remain outside the approved training set while the gap is investigated.