Reading Proficiency Assessment Using Finite-State Transducers

This paper describes a model designed to detect miscues of children's oral reading. The model is evaluated on a corpus of primary school children with and without reading difficulties in Dutch. This set contains real words and pseudo words reading tests. The automatic speech recognition task is achieved by training an End-to-End (E2E) model with phonemic targets. The encoder employs the Conformer architecture, and the decoder follows the Transformer decoder scheme. The automatic assessment relies on modeling a lexicon at a phonetic level using a Weighted Finite State Transducer (WFST) that models the pronunciation lexicon. The proposed WFST construction accommodates all the pronunciations defined by the lexicon for any given word, allowing the assessment to handle multiple pronunciations. The experiments show that the accuracy of this setup outperforms previous results obtained on the same evaluation set.