How we collect speech data with consent

Speech models are made of people. Every hour of audio behind a voice model is an hour of somebody's time, somebody's accent, somebody's way of laughing halfway through a sentence. For African languages, a lot of the speech data that exists today was scraped, collected without clear permission, or gathered once and never paid for properly.

We do not want to build Waridi that way. This post explains how our collection programme works, what we ask of speakers, and what we promise in return.

Consent starts in the speaker's language

A consent form written in dense English legal prose is not consent for most people we record. So every part of our process is available in plain Swahili first, with English as a second option.

Before anyone records a single sentence, a coordinator walks them through the project out loud. Then the speaker reads or listens to a short summary and answers a few check questions. We want to know they understood, not just that they signed.

The summary opens with something like this:

Tunakuomba urekodi sauti yako ili tufundishe kompyuta kuelewa na kuzungumza Kiswahili. Unaweza kukataa au kujiondoa wakati wowote.

In English: "We are asking you to record your voice so we can teach computers to understand and speak Swahili. You can refuse or withdraw at any time."

What speakers agree to

We keep the agreement narrow and specific. A speaker agrees to:

• record prompted sentences, short conversations or read passages, for a session of a stated length;
• let Waridi use those recordings to train and evaluate speech recognition and speech synthesis models;
• let us keep a small amount of metadata (described below) to make the dataset balanced.

A speaker does not agree, by default, to have their voice cloned or offered as a named voice in Waridi Sauti. That is a separate, explicit agreement with its own terms and its own payment. We cover voice cloning in detail in Voice cloning and consent.

Paying speakers fairly and promptly

Recording is work. It takes concentration, a quiet room and patience with repeated takes. We pay for it like work.

Our rules for payment are simple:

1. The rate is stated before the session starts. No surprises, no "we will see".
2. Payment is per session, not per accepted clip. If our tooling rejects a recording because of a technical fault, that is our problem, not the speaker's.
3. Payment goes out quickly, by mobile money. Our goal is to pay within one working day of a completed session, directly to M-Pesa, Tigo Pesa, Airtel Money or similar, so speakers do not need a bank account or a trip to an office.
4. Travel and data costs are covered. If we ask someone to come to a venue or use their own mobile data, we reimburse it.

We set rates with reference to local wages for skilled work of similar length, and we review them as the programme grows. If speakers tell us a rate feels unfair, we listen.

The right to withdraw

Withdrawal has to be real, or the word "consent" means very little.

Any speaker can withdraw by contacting their coordinator or writing to hello@waridi.io. When they do, we:

• remove their recordings from our active datasets;
• exclude them from all future training runs;
• confirm the removal back to them in writing.

We are honest about one limit: a model that has already been trained cannot easily "forget" a single person's audio. What we commit to is that withdrawn recordings are never used in any new training or evaluation from the moment we receive the request. Withdrawal never affects payment for sessions already completed.

Collecting as little as we can

We need some metadata to build balanced datasets. A Swahili model trained only on young men from Dar es Salaam will fail everyone else. But we keep this to the minimum:

• broad age band (for example 18–29, 30–44), never a date of birth;
• gender, if the speaker chooses to share it;
• region where they grew up and region where they live now;
• languages they speak at home.

We do not store national ID numbers against recordings, and we separate payment details from the audio dataset. Recordings are linked to a random speaker code, not a name. Access to the mapping between codes and people is restricted to a small number of staff with a clear need.

This approach follows the principles in Tanzania's Personal Data Protection Act, 2022: collect for a stated purpose, collect only what is necessary, keep it secure, and respect the rights of the person the data is about. Our privacy policy and security page describe the wider controls.

No children

We do not record anyone under 18. Children's voices matter for future products, especially in education, but the ethical bar for collecting them is much higher, and we are not ready to meet it. If we ever do this, it will be through a separate programme designed with parents, schools and child-protection specialists, announced publicly first.

Community review

We do not want to decide alone what is fair. Before scaling collection in a new region, we share our consent materials, payment rates and prompt lists with local community members and ask them to challenge us. Prompts that are awkward, offensive or culturally off get removed. Instructions that confuse people get rewritten.

This also improves the data. Prompts written with local input sound like how people actually talk, which is exactly what we want the models to learn.

What we are aiming for

Our long-term edge is high-quality, ethically sourced African speech data. That only holds if the people who contribute their voices would happily do it again and tell their friends to join. That is the standard we are holding ourselves to.

If you have questions about our collection programme, or you want to take part, write to us at hello@waridi.io.