Your reading path
Start with the right questions.
From principle to practice
An illustrative voice-data processing matrix
Open the worked exampleStart with the actual processing, not a compliance label
A statement such as “GDPR-ready audio” is too broad to tell an engineering team what it may do. Write down the intended operations: receive original recordings, prepare transcripts, train a recognizer, share examples with an annotation partner, distribute weights, or operate a voice product. The people receiving the material and the information exposed can differ at each stage.
Also establish why GDPR might apply. The answer is not simply the location of a storage bucket or the incorporation address of the buyer. Ask the responsible privacy reviewer to assess the organizations, people and activities involved. Record the conclusion and its scope instead of treating a commercial vendor’s location as an exemption.
Separate identifiable speech from biometric identification
GDPR Articles 4 and 9 distinguish personal data and biometric processing for unique identification. Voice recordings may identify a person through their voice or the conversation’s content; they are not automatically special-category biometric data merely because they contain speech.
For a practical inventory, record whether names, contact details, location references, sensitive discussion topics or identity-related voice features are retained. A transcription project and a speaker-authentication project can use similar files for different purposes. Do not carry a prior assessment across those uses without reviewing the change.
Document the selected basis and its limits
GDPR Article 6 lists several lawful bases. A written copyright license is not itself a substitute for identifying the appropriate personal-data basis. Where consent is used, the actual information provided to participants and the agreed purpose matter; do not describe every source agreement as consent just because it has a signature.
The EDPB’s Opinion 28/2024 discusses necessity and balancing when legitimate interest is proposed for AI processing. That is an assessment to substantiate, not a shortcut that a dataset invoice establishes. Give reviewers the source context and realistic alternatives, including a smaller collection or less identifying representation.
An illustrative voice-data processing matrix
The following is an authored example for a proposed customer-support recognizer. It describes questions to answer, not an approved collection or a legal conclusion. Keeping the rows separate reveals where a source proposal is incomplete.
- Operation
- Source intake
- Information to record
- Recording source, dates, participant information and original purpose
- Question before proceeding
- What evidence supports this proposed reuse?
- Operation
- Transcript preparation
- Information to record
- Identifiers in text, annotation partner, access scope
- Question before proceeding
- Can unnecessary identifying details be excluded without damaging the task?
- Operation
- ASR fine-tuning
- Information to record
- Selected segments, stable IDs, split groups, training location
- Question before proceeding
- Which assessment covers this purpose and these recipients?
- Operation
- Model distribution
- Information to record
- Weights or hosted access, intended recipients and documented risks
- Question before proceeding
- Does the review cover distribution as well as internal training?
- Operation
- Retention and correction
- Information to record
- Source versions, derivatives, contact route and operational owner
- Question before proceeding
- How will a valid request be identified and acted on?
Explain what the matrix changes
Suppose an archive is cleared for internal recognition research but an annotation subcontractor has not been assessed. The project has a specific unresolved dependency; it is not useful to call the entire archive either “compliant” or “noncompliant.” The next action is to resolve that recipient and handling scope before making the transfer.
Likewise, a team may need technical features such as speaker grouping to keep evaluation independent without needing public speaker names. Store the minimal project identifiers and keep any necessary identity mapping separately restricted. This is an operational design choice to discuss with the reviewer, not proof that the audio has become anonymous.
Anonymisation and withdrawal need realistic engineering plans
The EDPB’s AI opinion treats model anonymity as a case-specific question. Replacing a name with a speaker code is therefore not sufficient evidence that every relevant risk has disappeared. Document the transformation and the residual information rather than promising anonymity based on a filename convention.
Create a lineage map from source recording to segments, transcripts and released dataset versions. A correction or deletion process can then locate affected artifacts. Do not promise that removing a source file automatically removes its influence from every existing trained model. The actual response depends on the applicable rights, circumstances and feasible, reviewed procedures.
Keep AI Act obligations separate
The EU AI Act contains distinct requirements for different roles and systems. Its general-purpose model provisions include copyright-policy and training-content-summary obligations. A speech-data purchase does not by itself establish which obligations apply to a buyer or satisfy them.
Ask the project reviewer to identify the organization’s actual role, applicable provisions and timing. From the data side, retain evidence that can support that review: source descriptions, preparation records, limitations and version history. Avoid selling a generic documentation package as a universal AI Act compliance certificate.
What to request before receiving a sample
Begin with a high-level source description and a proposed handling plan. A useful review packet exposes open issues without distributing private participant records to everyone evaluating a purchase. Redacted examples and controlled review may be more appropriate than attachments in an initial email.
- Identify the project purpose, responsible organizations and proposed recipients.
- State which personal or sensitive information the task needs and why.
- Describe the source, participant information and supporting documents actually available.
- Specify retention, access, transfers, security review and a contact for rights requests.
- Mark unknowns and assign an owner before source access or expansion.
Questions buyers ask
Sources & research notes
Public documentation reviewed 2026-11-13. Provider descriptions and published terms are attributed below. Published facts are attributed to their sources. Worked files are labeled as illustrative or as an inspected editorial source; neither implies commercial dataset clearance.
- Regulation (EU) 2016/679, GDPREuropean Union · Reviewed 2026-11-13
Articles 4, 6 and 9 distinguish personal data, lawful bases and special-category processing, including biometric identification. Applicability depends on facts and jurisdiction. Voice is not automatically special-category biometric data merely because it is a recording.
- Opinion 28/2024 on personal data in AI modelsEuropean Data Protection Board · Reviewed 2026-11-13
Model anonymity is assessed case by case; legitimate interest requires purpose, necessity and balancing assessments. Opinion does not make every AI model anonymous or authorize every AI-training project. Rights and other GDPR duties remain.
- Regulation (EU) 2024/1689, consolidated AI ActEuropean Union · Reviewed 2026-11-13
Article 53 establishes distinct documentation, copyright-policy and training-content-summary obligations for general-purpose model providers. Do not apply every GPAI or high-risk obligation to every speech-data buyer. Scope, exceptions and transitional application need project review.
