Backed by Y Combinator

The first AI that speaks
oral languages natively

The first AI-powered app to teach Mandinka, Susu, and Pular through voice. No textbooks. No flashcards. Just you and the language, the way it was meant to be learned.

53M+
Mandinka + Susu + Pular speakers
0
Apps to learn them
1
Until now

Duolingo can't teach you Mandinka.
We can.

Oral languages are learned through sound, not text. Kahn teaches the way your family would — by speaking with you.

01

Listen

Hear a word spoken in Mandinka by a real voice. See the image. Understand the meaning before you ever try to say it.

02

Speak

Say it back. The AI listens to your pronunciation and scores you in real time. No typing. Just your voice.

03

Learn from a teacher

An AI teacher guides you through each lesson. It adapts to your level — patient when you struggle, challenging when you're ready.

04

Have a conversation

"Which food is your favorite? Say it in Mandinka." Real dialogue, not multiple choice. The way language actually works.

The Kahn Model

Trained on hearing,
not scripture

The dominant African language models — Meta MMS, AfriHuBERT — were trained on Bible recordings under non-commercial licenses. You can't build a product on them. The vocabulary is religious. The cadence is liturgical. The audience is 85% Muslim.

The Kahn Model is trained from real speech — how people actually talk about food, family, greetings, and daily life. Recorded by native speakers. Commercially licensed. Owned by us.

Kahn means sound in Mandinka. It also means language, it could also refer to your neck. It depends on how you use it. Sound is language.

kahn_model.py

# Every other model

Sound Text English Text Sound

# Tone lost. Meaning lost. Culture lost.

# The Kahn Model

Sound Meaning Sound

# Tone preserved. Hearing = Language.

languages = [

"Mandinka", # 11M speakers — ZERO language apps

"Susu", # 2.5M speakers — ZERO language apps

"Pular", # 40M speakers — ZERO language apps

]

license = "commercial" # Not CC-BY-NC. Ours.

trained_on = "real speech" # Not Bible recordings.

Timing

Why now

Duolingo has 40+ languages. Not one from West Africa.

Mandinka, Susu, Pular — 53 million speakers combined. Zero apps to learn them. Zero courses. Nothing.

Existing AI models can't be used.

Meta and Google trained on Bible recordings under non-commercial licenses. The vocabulary is wrong. The license blocks products.

AI made this possible for the first time.

Real-time speech scoring, AI teachers that adapt to your level, voice synthesis that sounds natural. None of this existed two years ago.

Early Access

Be the first to hear Kahn speak

We're launching with Mandinka first. Susu and Pular coming soon. Join the waitlist to get early access.