Geko
Geko provides voice AI models for underserved languages, starting with Kazakh and Russian. Its speech-to-text model achieves 8.71% word error rate on Kazakh audio—five times better than Whisper—and supports mid-sentence language switching without a language parameter. The OpenAI-compatible API is live in production with four pilots in Kazakhstan, including two national telecom operators.
- Artificial Intelligence
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Major AI labs focus on high-resource languages, leaving languages like Kazakh with millions of speakers without functional voice AI. Existing speech models require users to select a language per recording, which fails when bilingual speakers switch between languages mid-sentence, resulting in high error rates and unusable voice experiences.
Solution
Geko builds and serves voice AI models specifically for underserved languages. The company records and licenses its own speech data, then trains models end-to-end on that data, enabling natural-sounding speech quality that scraped-data models cannot match. Geko offers two core models: Tokay-1.0 for text-to-speech (six Kazakh voices, WAV output at 24 kHz) and Seta-1.0 for speech-to-text (Kazakh and Russian in a single shared vocabulary, allowing seamless mid-sentence switching). The API is OpenAI-compatible, so developers can switch from other providers by changing only the base URL and model name. In production since July 2026, Geko's models are already deployed in four pilots in Kazakhstan, including two national telecom operators.
Target Audience
Primary customers are developers and enterprises building voice AI applications—such as call center automation, order status updates, and receptionist systems—in languages poorly served by global speech models, particularly in Kazakhstan and other Kazakh-speaking markets.
Features
- Text-to-speech model (Tokay-1.0) with six catalogued Kazakh voices (three female, three male) and sentence-by-sentence streaming
- Speech-to-text model (Seta-1.0) supporting Kazakh and Russian in one shared vocabulary with no language parameter, enabling mid-sentence and mid-word switching
- 8.71% word error rate on Kazakh audio, compared to Whisper's 44.95%, and 40× faster than real-time transcription
- OpenAI-compatible API endpoints for both directions, allowing drop-in migration without changing call sites
- Self-serve metered billing per character synthesized and per second transcribed, with GPU backends scaling to zero for cold starts
- Training on proprietary voice data recorded and licensed by Geko, with models available for custom languages on request