ROBUST UZBEK ASR AND TTS FOR DIALECTAL AND NOISY SETTINGS

Abstract

In this article we present a unified recipe for robust Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) for Uzbek under dialectal variation, code-switching, and real-world noise. We combine self-supervised pretraining, dialect-aware lexicons, multi-script text normalization, targeted augmentation, and test-time adaptation. On simulated and field recordings, the ASR reduces WER by large margins; the TTS maintains naturalness and intelligibility across accents and SNRs.