BUILDING AN OPEN UZBEK SPEECH-TO-TEXT STACK FOR LOW-RESOURCE ASR

Abstract

In this article we present an open, end-to-end speech-to-text stack for Uzbek that lowers the cost of building practical ASR in low-resource settings. The stack combines reproducible data pipelines, self-supervised pretraining, lightweight fine-tuning, and lexicon-free decoding. On public Uzbek splits, fine-tuned wav2vec 2.0 and Whisper outperform strong baselines. We release recipes, tokenizers, and evaluation scripts to enable fair, repeatable benchmarking and rapid local adaptation.