Застосування методу GUIDED DECODING у системах автоматичного розпізнавання мовлення для оцінювання якості читання вголос

Abstract

This paper investigates Guided Decoding in ASR systems for assessing reading quality in educational mobile apps. The method biases token probability distribution towards an expected text while preserving error detection capability. Two models were compared: Moonshine Tiny UK for Ukrainian and Whisper Tiny as a multilingual baseline, evaluated in classic and Guided Decoding modes on Ukrainian text. Results show Guided Decoding consistently reduces WER and CER, with optimal bias range of 5–10 providing the best trade-off between recognition accuracy and error detection capability