About Clarion
Clarion is a team in Tashkent, Uzbekistan, building speech recognition for the Uzbek language. It is available as an API for audio files, as realtime transcription and as a model you can run on your own server. We build it for call centers, media and developers.
Why we build it
General-purpose models handle Uzbek poorly, conversational speech most of all.
Tens of millions of people speak Uzbek, yet general-purpose speech recognition handles it poorly. On real conversations these models drift into neighbouring Turkic languages, drop whole phrases and invent text over silence.
Most open Uzbek data is read speech, while call centers and media work with conversations, interviews and broadcasts. So we train a dedicated model on conversational Uzbek and test it on the same kind of speech.
- 8.1%uzasr-large word error rate (WER) on our open benchmark of real conversational Uzbek
- 38.6%whisper-large-v3-turbo-uzbek, a Whisper model fine-tuned for Uzbek, on the same benchmark
- ×4.7fewer errors than Whisper fine-tuned for Uzbek
How we work
Numbers you can check, and a public record of what went wrong.
An open benchmark
We measure every model on 311 clips of spontaneous Uzbek speech from podcasts, interviews and multi-speaker conversations. The benchmark, the scoring script and the model weights are public on Hugging Face, so anyone can check our numbers.
Corrections in public
We withdrew release 0.2.0 a day after publishing it: its language model had seen part of the benchmark. The 0.2.1 post explains how we found the leak and gives the corrected numbers. Later a native speaker re-checked the benchmark reference by ear. We published both versions and stated that the method favours our latest model.
One model in the cloud and on your server
The self-hosted license includes the same model that runs our API and realtime transcription.
Audio is deleted after transcription
We use recordings only to transcribe them. Logs and metrics contain neither audio nor text.
Team
Rustamjon and Abbosbek studied together and have been building products together for six years. All three of us work in Tashkent.
Rustamjon Nuriddinov
Co-founder, AI/ML engineer
Data collection and preparation, model training, evaluation and every model release. Before Clarion he spent over a year as a founding engineer at PromptLab, a US startup, leading the core backend infrastructure team: AI agents and MCP servers.
Abbosbek Fayzullayev
Co-founder, backend engineer
Amir Rustamov
Co-founder, product designer
The web platform, the dashboard and the API documentation.
Contact
Questions about the API, pricing or running the model on your own server.
- rustam@clarion.uz
- Telegram
- Location
- Tashkent, Uzbekistan