USC — ISSAI's Open Uzbek Speech Corpus (105 Hours, 958 Speakers) for Voice Technology in Uzbek

Uzbek, spoken by tens of millions of people, had almost no open speech data, which left voice interfaces, captioning and screen-reader-style tools out of reach for many users. In 2021 a team from ISSAI (Nazarbayev University, Kazakhstan) and Uzbek collaborators released USC, an open-source corpus of 105 hours of transcribed read speech from 958 speakers, manually checked by native speakers.
The authors trained DNN-HMM and end-to-end recognisers on it and reported word error rates of 18.1% on the validation set and 17.4% on the test set. They describe these results as preliminary. The paper was accepted at SPECOM 2021, and the dataset, pre-trained models and training recipes are shared on GitHub.
Evidence is limited to technical benchmarks: no user study, deployment or adoption data is published, and the dataset licence is not stated on the project pages. Its value for digital inclusion is as enabling infrastructure for Uzbek-language voice tools, not as a measured service outcome.

Read the full analysis: https://issai.nu.edu.kz/2021/09/29/usc-an-open-source-uzbek-speech-corpus-and-initial-speech-recognition-experiments/