Voice Cloning From Public Records: A Governance Teardown

The Norwalk, Iowa case isn't a security breach. Nothing was stolen, and the outcome was positive. However, it's a valuable case study because it shows a records program in action and highlights potential pitfalls in similar projects. The controls that worked here were mostly informal. Your team shouldn't rely on informality.

What Happened

Robin Leaper joined the City of Norwalk's Parks and Recreation Department in 2018 and later became its director until 2025. After being diagnosed with ALS in 2023, she began losing her ability to speak. She approached Tai Lieu, the city's Marketing and Communications Specialist, to help her reclaim her voice.

Lieu reviewed hours of recordings from public meetings and community events, isolated samples of Robin speaking, and used ElevenLabs' voice-cloning technology to create a synthetic version of her voice. By typing into a text-to-speech program, Robin could communicate with family, friends, colleagues, and her community in a voice that was recognizably her own.

These recordings were initially created for routine public business. Their secondary use was unforeseen, but it worked because the recordings were retained, findable, and usable.

Timeline

• 2018: Leaper joins the Parks and Recreation Department. Recordings of public meetings and community events begin accumulating.
• 2023: Leaper is diagnosed with ALS and begins losing her ability to speak.
• After diagnosis: She asks Lieu for help. Lieu searches the recordings, isolates voice samples, and generates the synthetic voice with ElevenLabs.
• 2025: Leaper's tenure as director ends.
• July 2026: Leaper dies after a three-year battle with ALS.

The sequence is important. The records existed years before anyone imagined this use. Nobody scheduled them for voice cloning.

Which Controls Held, and Which Were Thin

Consider this a gap analysis. Some controls clearly worked, while others might be overlooked in less careful projects.

Held:

• Retention and availability. Recordings spanning several years were retained, accessible, and usable. Without that, there was nothing to sample.
• Consent. Robin gave explicit permission for her voice to be recreated for her own use.

Not evidenced, and where your risk sits:

• Authorization trail. A durable record of who approved the reuse, why the synthetic content was created, and which source recordings and tools were used.
• Provenance linkage. Metadata connecting the generated audio to its source recordings and how it was made.
• Control of the voice model. Who holds it, who may use it, and what happens to it after the individual's death. Robin's death makes this question concrete.
• Vendor terms. Whether an AI vendor may retain source recordings or use them to improve other products.
• Labeling. How audiences are informed that the audio is AI-generated, so it isn't mistaken for an original recording.
• Recordness. Whether generated output becomes a record when used in government business, and if so, who declares it.

These are the questions a governed project must answer. It's not a claim that Norwalk failed to answer them.

What the Standards Require

The Generally Accepted Recordkeeping Principles, in ARMA International's updated form, provide a framework for evaluating generative AI applied to audiovisual records. Four Principles map directly to this case.

Record Lifecycle Management and Availability. Selected audiovisual records must be preserved in usable formats with sufficient metadata and discoverability. Lieu could find and prepare voice samples because that condition was met. Check your own Records Control Schedule: can you locate a specific speaker across years of meeting recordings today?

Accountability and Transparency. Document who authorized the reuse, why the synthetic content was created, which source recordings and technology were used, and how audiences will be informed the audio is AI-generated. This is a documentation duty, not a courtesy.

Compliance and Protection. Consent doesn't end your obligations. The source recordings, voice data, and resulting model still need access controls, vendor terms, privacy protections, and restrictions on further use.

Trustworthiness. Original recordings must remain distinguishable from generated speech. Provenance information should tie the synthetic output to its sources and describe how it was created. Labeling lets listeners understand what they're hearing.

The management-systems logic of ISO 30300 supports the same direction: records must be controlled by defined policy, roles, and processes. If your program has no stated owner for AI-derived outputs, that's a gap under any framework you choose. If you preserve source recordings under an Open Archival Information System (ISO 14721) model, your Preservation Description Information is the natural home for provenance linking originals to derivatives.

One point deserves emphasis. A recording being publicly accessible is not blanket ethical authorization to use a person's voice or likeness for AI-generated content. Public access and permission to synthesize are separate questions.

Lessons and Action Items for Your Team

1. Write a synthetic-media intake rule. Require documented authorization before any record is used as training or source material for generative AI. Record the requester, purpose, approver, source records, and tool.
2. Add consent to your file. Capture who consented, to which uses, and for how long. Decide in advance what happens to consent and the model after the person's death.
3. Settle recordness before the output exists. Define when generated content becomes a record, who performs Record Declaration, and which schedule it falls under.
4. Build provenance links. Use metadata fields that point each generated asset to its source recordings, the tool, and the creation date.
5. Review vendor terms line by line. Confirm whether the vendor can retain source recordings or reuse them to improve other products. Get restrictions in writing.
6. Assign model custody. Name an owner for any voice model, with access controls and a defined disposition path under a Records Disposition Authority.
7. Label by default. Set a standard disclosure for AI-generated audio so it can't be mistaken for an original.
8. Appraise with discipline. The possibility of future AI reuse can inform appraisal. It should not become a reason to keep everything. Retain what has legal, evidentiary, historical, or cultural value, and keep it trustworthy, accessible, protected, and governed.

Your job isn't to predict the next unforeseen use. It's to ensure that when one arrives, the records are still there, still authentic, and governed well enough to support it.