The protocol underneath African language AI.

A tokenizer that cuts the token premium, a consented recordings network and the app members record with.

Pixel mosaic of the Nzinga cowrie in magenta, flame and Volt
FLORES+ · SWAHILI
1.35x
▼ 8.4% vs GPT-4o
Recording a Swahili note
0:42
Pixel mosaic blooming from pale pink to a Volt core
Your choice
SWAHILI
✓
Only me
Not paid
✓
Nzinga's models
Paid
✓
Nzinga and approved partners
Paid + share
Everyone
Paid

Works with, measured on and released through

Gemma 3
Llama 3.1
FLORES+
NTREX-128
Hugging Face
Expo
WHY NZINGA

For the same meaning, African languages need more tokens than English.Nzinga cuts that premium at the tokenizer, then builds the data to go further: recordings members choose to share, paid for every accepted minute.

Three products built on one protocol.

A tokenizer, a consented network and a phone app. Every recording keeps its speaker's choice from start to end.

Nzinga Tokenizer

Fewer tokens for African languages.

32,000 African word pieces added to Gemma 3's tokenizer. Every existing token keeps its ID, so English and French barely change.

Cost

At any per-token price, fewer tokens cost less per query.

Context

More African-language text fits in the same window.

Speed

Fewer tokens to generate, so replies come back sooner.

Swahili

A 128,000-token window reads 79,100 words of Swahili, against 71,900.

Pixel mosaic in pale pink with Volt cores at the edges
Token count panel for Swahili on FLORES+: Nzinga 36,583 tokens, GPT-4o 39,922, Gemma 4 44,433, 8.4% fewer than the best other
Raw tokens

Counted on the same sentences, against twelve other tokenizers, never on premiums alone.

Held-out

No test sentence shares a 13-word sequence with any of the training text.

Pixel mosaic in pale cyan with one magenta cluster
Recording panel for a Yoruba voice note: recorded, choice set, details sent, accepted, with a withdraw button
Per-speaker consent

Every voice in a recording makes its own choice.

Withdrawal

Withdraw any recording from its own page, at any time.

Nzinga Network

Recordings that stay with their speakers.

Invite-only membership, a choice on every recording, pay per accepted minute and hashed contacts.

Invites

Three invites to start, and one more per accepted hour.

Pay

Paid per accepted minute, with a share of licensing revenue.

Contacts

Numbers scrambled on the phone; Nzinga keeps none of them.

Demo app

Opens in Expo Go: any number, code 482917, any invite code.

Or build on the API directly.

Python
TypeScript
cURL
lib/record.ts
1
import{ createLiveClient } from"./lib/api";
2
3
constapi = awaitcreateLiveClient(() => session.token);
4
5
constrec = awaitapi.createRecording({
6
kind: "voice_note",
7
language: "swh",
8
speakers: ["m_9f2c", "m_41ab"],
9
durationSeconds: 192,
10
});
11
12
for await(const r of (await api.myRecordings()).recordings) {
13
console.log(r.language, r.status);
14
}
294,145
tokens in the vocabulary
1.55x
average Tier 1 token premium
21 of 21
languages ahead on FLORES+

Consent you can check yourself.

Every consent names the exact text shown, by version and hash. Records are never edited, so history stays.

SHA-256 CONSENT HASH
APPEND-ONLY RECORDS
PER-SPEAKER
WITHDRAWABLE

Speakers

Each voice in a recording chooses for itself.

Hashed contacts

Numbers scrambled on the phone, never stored.

Versioned

Consent names its text by version and hash.

Invites

Members join only with an invite code.

Withdrawal

Withdraw any recording, or everything, any time.

Your choice

“Only me” never trains any model.

Pay that follows each minute you share.

Members

Paid

For every accepted minute

  • Rate set at launch
  • Your voice only
  • Three invites
  • Mobile money

Partners

Licence

Partner licensing, terms on request

  • Only consented, accepted recordings
  • Every speaker's choice checked
  • Language datasets
  • Revenue shared

Rates are set per market at launch. Members keep what they are paid, even after withdrawing.

Questions, answered.

The docs go deeper, or write to the team.

Who decides how my recordings are used?

You do, recording by recording: “Only me”, “Nzinga's models”, “Nzinga and approved partners” or “Everyone”. You can change or withdraw it any time.

Which models use the tokenizer?

Gemma 3 models today, and Llama 3.1 through a separate pack.

What counts as an accepted minute?

Each minute of your own voice that a reviewer accepts. Every choice except “Only me” is paid.

Can I withdraw a recording later?

Yes, any time. It leaves all future training and licensing. You keep what you were paid.

What happens to the contacts on my phone?

Each number is scrambled on your phone first. Nzinga keeps none of the codes.

How do I join Nzinga?

With an invite code from a member. The demo opens in Expo Go with code 482917.

Try it in your language.

Paste a sentence and see the tokens it needs.