STT Engine · Early access

An Uzbek-native
speech-to-text engine

The same engine that runs inside Bayonic VOCO — soon available as a standalone API. 98% accuracy on Uzbek on our internal test set, native UZ↔RU code-switching, your own vocabulary. Built for enterprise, deployable on-premise.

Request early access See it live in VOCO
98%
Accuracy on Uzbek · internal test
3
Languages
< 1.2s
Latency
On-Prem
Deployment
Why it matters

Built for the language global vendors left for last

Global STT vendors trained on English, Mandarin and Spanish. Uzbek — spoken by 35M+ people across Central Asia — was an afterthought, where it was supported at all. We built ours from scratch, on real contact-centre audio.

35M+
Native Uzbek speakers
Across Uzbekistan and the diaspora — and global models still treat the language as secondary.
2
Scripts supported
Latin and Cyrillic — both are in daily use, both are recognised natively.
UZ↔RU
Code-switching
People here change language mid-sentence. The model handles that on its own.
10k+
Hours of training audio
Real contact-centre audio from Uzbekistan.
Engine capabilities

What the API will deliver

When standalone access opens you get the full feature set already running inside Bayonic VOCO — exposed as a clean API.

Trilingual, natively
Uzbek (Latin & Cyrillic), Russian, English — one model, one endpoint, automatic language detection.
Code-switching
UZ↔RU inside one utterance, handled natively — no manual language flag per call.
Real-time streaming
WebSocket and gRPC under 1.2s latency. Suitable for live agent assist and IVR routing.
Speaker diarisation
Agent and customer separated even on a single mono channel — no special capture needed.
Custom vocabulary
Upload product names, brand terms and internal jargon, and the model spells them the way you do.
On-prem & cloud
Your choice of deployment. Air-gapped installs for banks and government; managed cloud for fast prototyping.
Roadmap

Where we are. Where we are going.

The engine is already in production — inside Bayonic VOCO. Standalone API access opens after the closed beta.

Shipped · 2024
v1 — embedded in VOCO
The first trilingual model trained and deployed.
Shipped · 2025
v2 — 98% accuracy, code-switching
A full retraining cycle on 10k+ hours of real contact-centre audio. Native UZ↔RU code-switching added; latency dropped below 1.2s.
3
In progress
Standalone API — closed beta
A small group of partners (banking, telecom, healthcare) gets API access for pilot integrations. Rate limits, SLAs and pricing tested.
4
Planned
General availability — public API
Self-serve access opens. REST and streaming endpoints. Per-second pricing, with on-prem licensing for regulated sectors.
5
Planned · 2027
v3 — domain models + Karakalpak, Tajik
Domain-tuned models (medical, legal, financial). Karakalpak and Tajik added for full Central Asian coverage.
Early access waitlist
Be first on the API

We are onboarding a small set of design partners for the closed beta. Banking, telecom, healthcare and government deployments get priority. Tell us about your use case.

Closed-beta priority
Founder-led onboarding
Custom vocabulary tuning
Launch pricing locked in