COMPARATIVE FINE TUNING STRATEGIES FOR UZBEK AUTOMATIC SPEECH RECOGNITION ACROSS MULTILINGUAL FOUNDATION MODELS

Abstract

This article presents a comparative analytical study of six model families relevant to Uzbek automatic speech recognition, namely Whisper, wav2vec 2.0, XLS R, MMS, SeamlessM4T, and Uzbek end to end systems trained on the USC corpus. The analysis is based on published academic evidence rather than a newly run laboratory benchmark. The central question is practical and linguistic at once. Which architecture is most suitable for Uzbek fine tuning under limited transcribed data, script variation, and morphologically dense word forms. The comparison integrates corpus scale, pretraining logic, adaptation behavior in low resource settings, language model compatibility, and risk zones specific to Uzbek speech and orthography