Automating Children’s Guided Reading Assessments with Automatic Speech Recognition

Assessing children’s reading skills is an important task for educators to keep track of a child’s progress. Guided oral reading assessments are a systematic task that areoften given to children to allow educators to provide feedback to children, and to identify areas of need. However, the requirement of a human listener to give feedback can create logistics issues in scaling assessment to larger groups. To this end, automatic speech recognition is a powerful tool that can be used to automate these types of oral reading assessments. We propose the development of a novel ASR system and metrics that are suitable for the assessment of reading in children. We plan to use transfer learning to fine-tune a pre-trained ASR system on the speech of children to improve performance within this subpopulation. We also plan to adapt ASR metrics to create a holistic measurement of reading literacy. We propose the integration of segmental and suprasegmental measures such as goodness of pronunciation and prosody-based features. With the model and metrics, we aim to create a new literacy tool for the automated assessment of guided oral reading in children. The development of an automated reading assessment tool can lower the resource burden on educators and increase the amount of feedback children receive when reading as well as providing a pathway for future child-oriented ASR applications. We will report on updates relevant to the development of child-specific ASR, in particular phoneme-recognition.

Thursday, May 16, 2024

9:30 – 10:00 AM ADT

Brian Diep
University of British Columbia

Dr. Bryan Gick
Professor, Department of Linguistics
University of British Columbia