Oye Pepa! — Voice corpus terms

AI Generated Human Curated

Version 1.0 · In effect from September 26, 2026 · Versión en español

These terms are in effect. Legal review has been requested and is pending. Any change follows Section 14: a new version may add protections for recordings already received, never remove them.

In plain words

Your voice stays yours. You lend it, free, to Pepa for one job, teaching a machine to hear “Oye Pepa!”, and you can take it back from every future use.

You keep: ownership of your recording and your voice, the right to withdraw, and the choice to be credited or not.
Recipients may: train, test, and study wake-word and keyword-spotting models, and publish those models under the model terms in Annex A.
Recipients may not: identify anyone, clone or imitate a voice, classify people by their voices, sell or charge for the recordings, or post the corpus publicly.
Access is by request, so a withdrawal can reach every copy.
Possession is not authority. Holding a copy never gives anyone more rights than the contributors granted.

This summary is for reading. The numbered sections are what bind.

1. Definitions

Contributor: a person who sent a Recording under these terms.
Recording: an audio file a Contributor sent.
Derived Artifact: anything made from a Recording: segments, format conversions, spectrograms, features, embeddings, augmented or noise-mixed copies, and cached training sets.
Corpus: all Recordings and Derived Artifacts, with their Contributor IDs and processing records.
Contributor ID: a random code assigned by the Steward, not derived from any personal data. The only link between an ID and an email address is a private table kept by the Steward.
Steward: PR Software Services LLC, San Juan, Puerto Rico, operator of ai4aging.org, which collects, keeps, and shares the Corpus.
Recipient: a person or institution granted a copy of all or part of the Corpus.
Trained Model: a machine-learning model whose parameters were materially trained or fine-tuned on the Corpus for the Purpose, and any modified version that keeps those parameters. A larger system that merely uses a Trained Model’s output is not itself a Trained Model.
Purpose: the permitted use in Section 3.

2. What Contributors confirm, keep, and grant

2.1 By sending a Recording, the Contributor confirms that they are 18 or older, that the voice in it is their own, and that it contains no other person’s voice as far as they know.

2.2 A Contributor aged 18, 19, or 20 also confirms that they understand these terms and what follows from agreeing to them.

2.3 A person under legal guardianship (tutela) may contribute only with the written consent of their guardian, sent with the Recording.

2.4 The Contributor grants the Steward a free, non-exclusive, revocable, worldwide licence to use the Recording for the Purpose, and to share it with Recipients bound by these terms. No payment is made or owed. This is a licence, not a transfer (cesión) of any right.

2.5 Each Contributor keeps all rights in their Recording and in their voice and image.

2.6 The Contributor may withdraw a Recording at any time under Section 6.

2.7 The Steward holds the Corpus in trust for its Contributors and may not give any Recipient broader rights than those in 2.4.

3. Purpose

3.1 The Corpus may be used only to train, evaluate, benchmark, and document wake-word and keyword-spotting models, and for research on those tasks.

3.2 Any other use needs new, specific consent from each affected Contributor.

3.3 The Corpus and the Recordings are not used to advertise or promote Pepa or any product, and no Contributor’s voice is ever played or published for that purpose.

4. Prohibited uses

No one may use the Corpus or a Trained Model to:

Identify, re-identify, or try to identify any Contributor, including by linking Recordings to other data.
Create, train, or operate a system whose purpose is to reproduce, synthesize, clone, convert, emulate, or otherwise generate speech intended to sound like a Contributor.
Build or train speaker recognition, speaker verification, or voiceprint systems.
Infer or classify, as a goal of the use, any Contributor’s age, sex, health, emotion, geographic origin, identity, or other personal characteristic.
Sell, rent, or license the Corpus or any Recording for a fee.
Surveil, profile, or make decisions about any person.

These limits are contractual. They apply in addition to any protection under Puerto Rico’s Ley del Derecho sobre la Propia Imagen (Ley 139-2011, as amended by Ley 163-2026) or other law.

5. Access

5.1 The Corpus is not posted publicly. A Recipient requests access from the Steward and agrees to these terms by replying “I agree” to an email that attaches the exact version of these terms.

5.2 The Steward logs every grant: who, when, which terms and Corpus version, and which subset.

5.3 A Recipient may not pass the Corpus on, except within its own team working on the Purpose under these terms, and may not charge anyone for access to it.

5.4 Recipients should process the Corpus on hardware they control. Any third-party service used must not keep the audio or train its own models on it.

6. Withdrawal (“retirar”)

6.1 A Contributor withdraws by replying to their original email with “retirar” or “withdraw”. No reason is needed.

6.2 Within 30 days the Steward removes the Recording and its Derived Artifacts from the Corpus, stops using them for any new training, and notifies every Recipient that received them.

6.3 Within 30 days of notice, each Recipient deletes every copy of the Recording and its Derived Artifacts, including features, embeddings, augmented copies, and cached training sets, and confirms deletion to the Steward.

6.4 A Trained Model already trained on a withdrawn Recording need not be retrained or deleted. Its next version must not use the Recording.

6.5 If a Recording turns out to include another person’s voice, or to come from someone under 18, the Steward deletes it and its Derived Artifacts as if withdrawn, and may ask the Contributor for a new Recording.

6.6 For a Contributor aged 18 to 20, a parent with legal authority (patria potestad) or a guardian may also request withdrawal, and it is handled the same way.

7. Trained Models

7.1 Pepa’s source code is under the MIT license. The MIT license does not apply to the Corpus, to any Recording, or to any Trained Model. Trained Models are published under the model terms in Annex A.

7.2 The intended output of a Trained Model is wake-word or keyword detection. The Steward will not knowingly publish a Trained Model designed or configured to reproduce, synthesize, convert, or imitate a Contributor’s voice.

7.3 Every published Trained Model names the Corpus version it was trained on and carries the terms in Annex A.

8. Credit

8.1 A Contributor is named in the Corpus credits only if they asked, by including their name with the Recording or later by email. Otherwise they are never listed publicly, not even by Contributor ID.

8.2 Institutions whose communities contributed are credited by name, with the number of Recordings each accounts for.

8.3 Suggested citation: Oye Pepa! Voice Corpus, version [number], ai4aging.org, [year].

9. Institutions

9.1 An institution whose members contributed Recordings may receive a copy of that subset, or of the whole Corpus, as a Recipient under Section 5.

9.2 An institution has the same rights and duties as any other Recipient, including honoring withdrawals under Section 6. It takes on no stewardship duties.

9.3 An institution’s own consent forms apply only to recordings it collects itself. They cannot widen use of Recordings collected under these terms.

9.4 Before announcing the collection to its members, an institution confirms in writing that taking part is voluntary and has no effect on grades, credit, enrollment, employment, or services.

10. Stewardship, succession, and discontinuation

10.1 The Steward keeps the Corpus on hardware it controls. The table linking Contributor IDs to email addresses is stored apart from the audio and used only for withdrawal, credit, and the notices in Sections 11 and 12.

10.2 The Corpus may pass to a successor only if the successor accepts these terms in writing. If the Steward can no longer act and no successor accepts within 90 days, the Corpus and the ID-to-email table are deleted and every Recipient is told to delete its copies.

10.3 The Steward may discontinue the Corpus at any time. It then deletes the Corpus and the ID-to-email table and tells every Recipient to delete its copies under Section 6.3. Trained Models already published remain available under Annex A.

11.1 Every Recording is logged with its arrival date, its Contributor ID, and the version of these terms in force when it arrived. The consent log, not the automatic reply, is the record of consent.

11.2 The Steward sends an automatic acknowledgment, in Spanish and English, in reply to submissions (at most once per day to a given sender), that confirms receipt and states:

the terms version, with a link to it;
the purpose: wake-word and keyword-spotting research only;
the duration: until withdrawn;
the territory: worldwide, for that purpose only;
the compensation: none;
the AI use: training detection models, with noise and speed augmentation for robustness; never cloning, synthesis, or voice imitation;
how to withdraw: reply “retirar” to the same address.

11.3 A Recording is governed by the terms version recorded for it, as limited by Section 14.

11.4 The consent log is kept for as long as the Recording is in use, plus one year. After a withdrawal, the Steward keeps only the Contributor ID, the date of withdrawal, and the date deletion was confirmed.

11.5 Recordings received before version 1.0, under the earlier wording of the instruction page, are governed by these terms, which only add protections. Those Contributors are notified and may withdraw at any time.

12. Security and breach notice

12.1 The ID-to-email table is stored encrypted on the Steward’s own hardware, readable only by the Steward, and is never sent to Recipients.

12.2 If the Steward learns that the table or the Corpus was accessed without authorization, it notifies the affected Contributors within 10 days and meets any legal duty that applies.

13. Transparency

13.1 For each Corpus version the Steward publishes the number of Recordings and Contributors, the processing steps (for example: converted to 16 kHz mono, split into single utterances, augmented only for training, no voiceprints stored), the number of withdrawals, and the number of Recipients.

13.2 Anyone may ask the Steward what a Recording is being used for and gets a straight answer.

14. Changes

14.1 These terms are versioned. Each change is added to a public changelog; earlier versions stay published unchanged.

14.2 A new version may add protections for existing Recordings. It may not remove any without new consent from each affected Contributor.

15. Language

These terms are published in Spanish and English. If the two differ, the Spanish version controls.

16. Termination, remedies, severability, law

16.1 A Recipient that breaks these terms loses access at once and must delete every copy of the Corpus.

16.2 Loss of access does not limit any other right or remedy available under these terms or applicable law.

16.3 If any part of these terms is found unenforceable, the rest still applies.

16.4 The Corpus is provided as is, without warranty of any kind.

16.5 These terms are governed by the laws of Puerto Rico, a U.S. territory. Any dispute goes to the Court of First Instance, San Juan Region.

Annex A — Model terms

These terms travel with every Trained Model.

A.1 You may use, copy, modify, and redistribute this model, free of charge.

A.2 You may not use this model, or anything derived from it, to identify people, to clone, synthesize, or imitate a voice, to recognize or verify speakers, to classify people by age, sex, health, emotion, origin, or identity, or to surveil or profile anyone.

A.3 When you redistribute the model, keep these terms with it and name the Oye Pepa! Voice Corpus version it was trained on.

A.4 Breaking A.2 ends your right to use the model.

A.5 The model is provided as is, without warranty of any kind.

Contact

Withdrawals: reply “retirar” to your original email, or write to oyepepa@ai4aging.org. Questions: editor@ai4aging.org.

Changelog

1.0 — September 26, 2026. First published version.