On the possibility of using pre-trained ASR-models to help assess oral reading exams
Keywords
Loading...
Authors
Issue Date
2024-08-31
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Dutch children’s reading skills have been declining for years. Oral reading skills in primary schools are often assessed using the three-minute-exam (DMT), a time-consuming one-on-one test where teachers manually judge word reading correctness. To reduce the workload, automatic speech recognition (ASR) could assist, but many ASR models struggle with children's speech. However, the DMT only requires correct/incorrect judgments, not perfect transcription.
We tested two ASR models, wav2vec2.0-CGN and faster-whisper-v2, using the Children’s Oral Reading Corpus (CHOREC). Faster-whisper-v2 outperformed wav2vec2.0 in accuracy (.89 vs. .69), F1-score (.58 vs. .39), and MCC (.54 vs. .37). However, the imbalance of the dataset may skew this result. Rule-based improvements enhanced performance more than similarity-based improvements. While ASR models show promise, they are not yet reliable enough for classroom use. Future research should focus on improving ASR performance and better utilizing diagnostic data to enhance children’s reading skills.
Description
Citation
Supervisor
Faculty
Faculteit der Letteren