# A similar voice is not permission

Permission evidence and clip acceptance answer different questions. A voice workflow and an isolated predicate trial using fictional data.

L00P.AI · 1.0 · 2026-09-26

https://www.l00p.ai/en/resources/series/podobny-glos-nie-jest-upowaznieniem/

## 01 / Two different questions

A system says: “this passage resembles person A's voice”. An editor asks: “may we use it to create new utterances?”. These are different questions. Recognition does not replace a document, and a document does not establish that the clip contains the right speaker.

This guide develops internal voice-use rules and clip-preparation documentation. It presents an acceptance method without samples, running clones or disclosing participant data. It does not determine rights to any particular recording.

## 02 / Start with scope

Separate file access, rights to use a recording, permission to synthesize new utterances and the basis for processing personal data. Owning software or a voice bank does not resolve these questions. A publicly played programme is not a document authorizing unrestricted synthesis either.

Before preparing a collection, record its purpose, permitted channels, duration, audience and responsible person. If a representative acts, check their authority. “Consent confirmed” in code points towards evidence; it is not the evidence itself. Keep private documents out of a publication's public record.

The proposed working card contains a reference to authorization evidence, permitted use, restrictions, review date, reviewer and a way to stop subsequent uses. A changed purpose, such as moving from training narration to advertising, calls for a renewed scope check.

## 03 / Recognition also needs assessment

The [GDPR](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679) distinguishes personal data from biometric data resulting from specific technical processing. Article 9 covers, among other things, biometrics used for unique identification. Not every voice recording automatically falls into this category; purpose and processing matter.

Avoid both “every voice is a special category” and “recognition needs no basis”. Explicit consent is one Article 9 condition, not the only one listed. Establish the basis, information duties and safeguards for the actual process. This card does not replace that assessment.

## 04 / Accept every clip separately

A speaker label assigned to a long programme does not guarantee a clean short passage. It may contain another speaker, an announcer, an archival quotation or overlapping voices. Preserve the original; make cuts in a copy with a time map.

For each candidate, record the source identifier, its checksum, IN/OUT boundaries and listening result. Automated voice comparison remains supporting evidence, recorded with the tool version and test conditions. We provide no universal matching threshold.

![Two separate checks: permission scope and clip acceptance. Missing evidence stops work. Both precede synthesis, content review and AI disclosure.](https://www.l00p.ai/wydawnictwo/podobny-glos-nie-jest-upowaznieniem/dwie-kontrole-en-v1.svg)

Original method diagram, Codex / L00P.AI. A voice match does not establish permission; a document does not establish clip content.

Hold uncertain clips for clarification. A target number of minutes must not force acceptance. After assembly, listen across cuts and compare the collection with the accepted list. A clip map leads back to the source; it does not establish permission to use it.

## 05 / What we checked in code

We read historical workflow documentation and the available clip-checking code. We locally executed its isolated acceptance predicate on four fictional records. We ran no recognition, accessed no voice bank and processed no audio.

A matching identifier and high score returned a positive result even when a separate status field said “unknown”. A different identifier and a missing score returned negative results. The predicate did not check authorization evidence. This describes that code fragment, not proof that the complete system permits publication.

The trial shows why an “OK” field name can mislead. Know precisely what was checked. [No alert does not confirm identity either](/en/resources/series/przedstawienie-pomaga-ale-nie-potwierdza-tozsamosci/). Four predicate responses are not a measurement of speaker-recognition accuracy.

## 06 / Synthesis must be identifiable as synthesis

Do not present a newly synthesized utterance as an authentic recording of the person. Record the input text, tool version, reference-material provenance and acceptance result. Check words, numbers, omissions and the context in which the message will appear.

The [European Commission explains Article 50 AI Act obligations](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act): deepfake disclosure should be clear and made no later than the audience's first exposure. A machine-readable mark alone does not replace information for people. Evidently artistic or fictional works have particular disclosure arrangements; this is not a blanket exemption from labelling.

Our practical rule is to make the disclosure understandable outside the original page layout, including audio-only listening. An end credit alone does not automatically prove compliance with every obligation. An AI label does not grant rights to someone else's voice.

## 07 / Stopping is part of the process

Before use, establish who can stop synthesis and publication, where copies are held and who can access references. Separate stopping new operations from deleting data and withdrawing an existing publication. The effects of changed or withdrawn authorization require assessment under its basis; do not promise remote deletion of copies downloaded by others.

A process description and intended use are enough to begin a deployment discussion. We can help prepare an authorization card, clip inventory and acceptance criteria. Private documents and recordings need not appear in a public presentation.

## 08 / Provenance and limits

Sources include an internal consent decision, a draft rights note, the opening parts of session documentation and the complete clip-checking script. Data-protection and transparency claims were checked against the linked official sources on 26 September 2026. Private people, identifiers, samples and historical model results were omitted.

Codex prepared the PL/EN text and original SVG and performed the described predicate trial. Review is by the author, without an independent second model. No acoustic test or authorization audit of particular people was performed. Changes to law, use scope or the checker, and discovered errors, trigger another review. A human may withdraw the publication.