Voice AI modelsfor underserved languages
Real-time speech in the languages the big labs skip, trained on voice we record ourselves
Meet Tokay-1.0
and Seta-1.0
Tokay speaks, Seta listens. Text to speech and speech to text on one API key, fast enough for a live call.
Speech,both directions
Geko builds voice AI models for the languages global speech models serve badly. Tokay-1.0 turns text into natural speech, Seta-1.0 turns speech into text, and both are trained end to end on voice data we record ourselves rather than scrape off the internet.
They run behind one HTTP API on one key, with a typed TypeScript SDK on npm, a command line tool, and OpenAI-compatible endpoints, so an existing OpenAI or Whisper client moves across by changing its base URL.
- Text to speech
- Tokay-1.0, six Kazakh voices, WAV at 24 kHz, sentence-level streaming
- Speech to text
- Seta-1.0, Kazakh and Russian in one model, including mid-sentence switching
- Accuracy
- 8.71% word error rate, where Whisper scores 44.95% on the same audio
- Integration
- HTTP API, TypeScript SDK on npm, a CLI, and OpenAI-compatible endpoints
- Billing
- Metered per character synthesized and per second transcribed, with free credits to start
- Guides
- Voice agents, and server-side integration for Next.js and Express
- Next
- More languages on the same endpoints
Hear it beforeyou sign up
Two of the six Kazakh voices, straight out of the API. These are the response bytes with nothing done to them — no editing, no mastering, no format conversion. Download either file and check.
Six voices ship today, three female and three male.
The voice catalogue →From nothing tospeech in threescreens
Geko is a self-serve API with a console in front of it. You create a key, hear a voice in the playground, then copy the same call into your own code and watch the usage land. No sales call in the way.
- 01
Create a key
Sign in, name your organisation, create a key. One key covers both models and both directions, and every new organisation starts with $1.00 of credits.
Open full size, 3024 × 1722 ↗ - 02
Speak a sentence in the playground
Type Kazakh, pick one of the six voices, set speed and quality, and listen. The cost of the request is priced in front of you before you run it.
Open full size, 3024 × 1724 ↗ - 03
Move it into your code
The same request over HTTP, the typed SDK or the CLI. Spend, characters synthesized and seconds transcribed land in the overview as you go.
Open full size, 3024 × 1720 ↗
npm install @gekoai/sdk
const geko = new Geko({ apiKey: process.env.GEKO_API_KEY });
const audio = await geko.tts.create({
text: "Сәлеметсіз бе!",
voice: "Aigerim",
});Three people,two models,one API
Geko was founded in July 2026 in Astana. We record the voice data, train the models and run the API ourselves, which is the only reason a company this size can serve languages the large labs have not got round to.
Need a voice that doesn’t exist yet?
Tell us the language, we record the speakers and train the model