For speech-language pathologists

How does consent work when children’s speech is recorded for research or AI?

Published · Updated

The short answer

In US research covered by federal human-subjects rules, a child takes part only with a parent’s permission and, when the child is able, the child’s own assent, under procedures approved by an institutional review board (IRB). Assent means active agreement: not objecting isn’t enough, and a capable child’s “no” generally prevails. Families should be told how recordings will be stored, shared and reused. Voices are hard to make anonymous, so it’s worth asking whether recordings could train AI and what happens on withdrawal.

Which rules apply

Parental permission and child assent

Permission is a parent’s or guardian’s agreement to their child taking part. Assent is the child’s affirmative agreement, and mere failure to object shouldn’t be taken as assent (45 CFR 46.402). In OHRP’s words, the child must actively show willingness rather than just comply (OHRP).

The IRB decides whether children can assent, considering their ages, maturity and psychological state. It can make that judgment for the whole study or child by child, and assent isn’t required if a child’s capability is too limited to consult them (45 CFR 46.408). The American Academy of Pediatrics describes permission and assent as an active process, with children’s opinions included more as they mature (Katz et al., 2016).

In practice, for a recording task with a young child, that can mean showing the microphone, explaining in a sentence what they’ll say and for how long, and treating reluctance or wanting to stop as a no.

Telling families how recordings are stored, shared and destroyed

The consent process must describe how far records identifying the participant will be kept confidential, and must say either that identifiers might be removed and the information used in future research without further consent, or that it won’t be used that way even with identifiers removed (45 CFR 46.116).

If researchers ask for broad consent to store and reuse identifiable data, they must describe the types of research it may be used for, what information may be used and shared, the kinds of institutions that might use it, and how long it may be kept, which can be indefinitely (45 CFR 46.116).

ASHA’s ethics guidance lists precautions such as storing research records securely, limiting access, removing or coding identifying information, and getting written consent from a parent before sharing findings that include video or audio recordings that might identify a child (ASHA).

Why voices are hard to de-identify

A voice is an identifier in its own right. HIPAA’s de-identification standard lists voice prints and full-face photographs among the identifiers that must be removed (45 CFR 164.514), and the updated COPPA Rule counts voiceprints as biometric identifiers (FTC, 2025). Speech recordings can be used for purposes ranging from health profiling to recognizing who is speaking (Nautsch et al., 2019).

Tools that disguise voices are still being tested. In a 2025 challenge, the best attack systems cut the error rate for matching anonymized speech to the right speaker by 25 to 44%, relative to the standard test attacker (Tomashenko et al., 2025). Disguising a voice doesn’t guarantee it can’t be linked back. Recordings can also contain names, schools or other details said aloud.

This matters because research rules allow de-identified data to be reused without new consent if the consent form said so (45 CFR 46.116). Ask exactly what “de-identified” will mean for audio.

Withdrawal, secondary use and AI training

Questions parents can ask before saying yes

For SLPs: if you collect recordings for a study or a product, or help recruit families, check with an IRB before you start. For apps and devices that record children outside research, see children’s voice data privacy.

This is general information, not legal advice.

Disclosure

Saylo, which publishes this page, is a speech-practice wearable in development. It hasn’t been tested with children yet. How Saylo plans to keep any research separate and consented is on its privacy and safety page.

Questions SLPs ask

Can a child refuse to take part in research if their parent agreed?

Yes, generally. If the IRB requires assent and the child is able to give it, the child’s “no” prevails, except in rare studies offering an important direct benefit available only through the research.

Does my child have to sign a consent form?

Not necessarily. US federal rules don’t require written assent; the IRB decides how assent is recorded, and for young children who can’t read it may simply be noted.

Can I withdraw my child’s recordings from a study?

Your child can stop taking part at any time. Under OHRP guidance, researchers may keep data already collected but can choose to delete it, so ask before you agree.

Can research recordings of children be used to train AI?

It depends on what the consent form says. De-identified data may be reused without new consent if families were told, and online services covered by COPPA need separate parental consent to share children’s data for AI training.

Is an anonymized voice recording really anonymous?

Not always. Voice prints count as identifiers under HIPAA, and research on voice anonymization shows stronger attacks can partly undo it.

Sources

Join the Early Adopter Program