Partner with us
All projects
Speech technology2026

Karakalpak speech recognition

A Karakalpak speech-to-text model. The word error rate is not yet at release quality, so we are holding the launch and fixing the data instead of shipping something that mishears people.

02 / Speech · research
In researchNot released
34%
Current WER
15%
Release target
38h
Training data
2026
Started
Project brief
PartnerCenter research
StatusIn research
TeamCenter speech team
Stack
PyTorchASR architecturesData collectionEvaluation with native speakers

The challenge

Recognition is harder than synthesis for a low-resource language: it needs many speakers, many recording conditions and many dialect variants. Our current error rate is too high for the uses people would immediately put it to — transcription of meetings, appeals and public records.

Our approach

01

Diagnose, do not patch

Error analysis by speaker, dialect and recording condition instead of chasing a single benchmark number.

02

Grow the corpus where it fails

Targeted recording of the speaker groups and conditions the model handles worst.

03

Release when it is honest

It joins the karakalpakvoice.uz API only when the error rate is low enough for real transcription work.

Note

We would rather publish late than publish a model that mishears the language. This page will be updated when the error rate is release-worthy.

More projects