Voice cloning and consent

A person's voice is part of who they are. Family recognises it on the phone. Customers trust it on a radio advert. Listeners follow it through a podcast. That is exactly why voice cloning is useful, and exactly why it can do real harm.

Waridi Sauti will let creators and businesses clone a voice through POST /v1/voices/clone. A narrator could produce audiobooks in their own voice without spending weeks in a studio. A small business owner could record a greeting once and update it in seconds. But the same technology could put words in the mouth of someone who never said them.

Our position is simple: a voice is only cloned when the person who owns it has clearly agreed, and they can take that agreement back. This post explains how we are building that into the product, not just into a policy page.

Consent you can hear

Ticking a box saying "I have permission" is not enough. Anyone can tick a box.

Instead, every cloning request must include a consent recording: the voice owner speaking a short statement, in their own voice, saying who they are and what they agree to. The statement is generated fresh for each request, so an old recording cannot be reused. In Swahili it reads something like:

Mimi ni Amina. Ninakubali sauti yangu itumike kutengeneza sauti ya kidijitali kwenye Waridi. Msimbo wangu ni 4 7 2 9.

In English: "I am Amina. I agree for my voice to be used to create a digital voice on Waridi. My code is 4 7 2 9." (The name here is only an example.)

We then check that the voice in the consent recording matches the voice in the training samples. If it does not match, the request stops. If the code is wrong or missing, the request stops.

What the request looks like

At the API level, consent is not optional metadata. It is a required part of the request:

curl https://api.waridi.io/v1/voices/clone \
  -H "Authorization: Bearer $WARIDI_API_KEY" \
  -F "name=greeting-voice" \
  -F "samples[]=@sample-01.wav" \
  -F "samples[]=@sample-02.wav" \
  -F "consent_recording=@consent.wav" \
  -F "consent_challenge_id=chl_..."

The full flow, including how to fetch a consent challenge, is in the voice cloning docs.

Identity checks

For personal use in Waridi Studio, the account holder cloning their own voice is the voice owner, and the consent recording is the main check.

When a business wants to clone someone else's voice, for example a presenter, a voice actor or a company founder, we require more:

• the voice owner verifies their identity separately from the account holder;
• the voice owner records the consent statement themselves;
• the account holder confirms the intended use (adverts, IVR prompts, narration and so on).

We will review high-risk requests manually, especially at the start of early access. That is slower. We think slower is right here.

Revocation

Consent that cannot be withdrawn is not really consent. Any voice owner can revoke a cloned voice at any time, even if someone else's account created it.

When a voice is revoked:

1. the voice stops working in the API and in Waridi Studio immediately;
2. new audio can no longer be generated with it;
3. the underlying voice model and samples are scheduled for deletion;
4. the account that created it is notified.

Audio already generated and published elsewhere cannot be recalled by us. That is one reason we care about provenance.

Watermarking and provenance

Our goal is for every piece of audio generated by Waridi Sauti to carry an inaudible watermark, so that we, and eventually others, can check whether a clip was produced by our system. We are also following open provenance standards for media and plan to attach provenance information where the format allows.

To be clear: this is a goal we are working toward, not something we claim is solved. Watermarks can be weakened by heavy compression, phone lines and re-recording. We will publish what our watermarking can and cannot detect once we have tested it properly.

What is not allowed

Our acceptable-use rules for cloned voices are strict. You may not use a Waridi voice to:

• impersonate a real person without their consent, including family members, officials or company staff;
• commit fraud, such as fake calls asking for mobile-money transfers, PINs or one-time codes;
• deceive voters or spread false political statements in someone's voice;
• create sexual or abusive content in a real person's voice;
• harass, threaten or defame anyone.

Breaking these rules means losing access, and where there is a risk of harm we will cooperate with the relevant authorities.

Public figures

Politicians, religious leaders, musicians and broadcasters are especially likely targets for fake audio. We will not clone the voice of a well-known public figure unless that person goes through the same consent and identity process as everyone else. Being famous does not make a voice public property.

We are also building checks that flag requests resembling well-known voices for manual review, even when the request claims consent.

Reporting abuse

If you believe your voice, or someone else's, has been cloned or misused through Waridi, write to hello@waridi.io with as much detail as you can: where you heard it, when, and a copy of the audio if possible. We treat these reports as urgent. If the voice was created on Waridi, we can disable it while we investigate.

Why we bother

Voice cloning for African languages is coming, whether we build it or someone else does. We would rather build it in a way people in Dar es Salaam, Nairobi and beyond can trust. Read more about our wider approach on the security page and in How we collect speech data with consent.