# Oye Pepa! — lend your voice

<span class="a4a-badge ai-generated">AI Generated</span>&nbsp;
<span class="a4a-badge human-curated">Human Curated</span>

Pepa is a local voice assistant — no cloud — for older adults living in their own homes. It wakes up when someone says **"Oye Pepa!"**, and to recognize that phrase it needs to hear it in many different voices. The current model was trained on synthetic voices and makes too many mistakes. Your voice fixes that.

It takes a minute.

## How to take part

1. Open the voice memo app on your phone.
2. Record **a single clip** saying **"Oye Pepa!" five times**, with a pause of a couple of seconds between each one.
3. Send it as an attachment to **oyepepa@ai4aging.org**, exactly as it comes off the phone. No need to convert or edit it.

[Send my recording](mailto:oyepepa@ai4aging.org?subject=Oye%20Pepa)

## Getting a good recording

- Speak the way you normally speak, as if you were really asking Pepa for something.
- Vary it a little: once faster, once slower, once as a question, once with the phone farther away.
- A reasonably quiet room is enough. It doesn't need to be studio silence.
- Every accent helps. The phrase is Spanish, but English speakers are welcome — Pepa lives in a bilingual house.

## Who can take part?

Anyone 18 or older. Older voices are especially valuable, because they're the ones Pepa most needs to understand. If you have parents or grandparents willing to record, their clip is gold - help them send it from their phone using their email address.

## Your recording, your rights

- **Taking part is voluntary.** Sending the recording is your consent to its use in an open dataset for training the wake word.
- **It's identified by an ID, not your name.** The audio and the ID are the only things that enter the dataset. Your email address is kept separately, only so a withdrawal request can be honored.
- **You can withdraw it.** Reply to the same email with the word **"retirar"** (or "withdraw") and the recording comes out of the dataset and out of all future training. A model already trained on it can't be "un-trained," but the next version won't use it.
- **Credit, if you want it.** If you'd like to appear in the dataset credits, include your name in the message. If you don't, you won't.
- **It isn't used for anything else.** No identifying people, no voice cloning, no selling data.
- **[Oye Pepa! — Voice corpus terms](https://starlight-page-actions.dlcastillop.com/general/oyepepa-terms-en/)** Applicable legal terms.

## What happens next

With the recordings received, I train a new wake-word model and document the whole process here: how many voices, how they were processed, how detection improved. The result is public, like the rest of the project.

Questions? Write to [editor@ai4aging.org](mailto:editor@ai4aging.org).