at Boulder My Science Tutor (MyST) is an intelligent tutoring system designed to improve science learning by elementary school students through conversational dialogs with a virtual science tutor in an interactive multimedia environment. Marni, a lifelike 3-D character, engages individual students in spoken dialogs following classroom investigations using the kit-based Full Option Science System program. MyST attempts to elicit self-expression from students; process their spoken explanations to assess understanding; and scaffold learning by asking open-ended questions accompanied by illustrations, animations, or interactive simulations related to the science concepts being learned. MyST uses automatic speech recognition, natural language processing, and dialog-modeling technologies to interpret student responses and manage the dialog. Sixteen 20-min tutorials were developed for each of 4 areas of science taught in 3rd, 4th, and 5th grades. During summative evaluation of the program, students received one-on-one tutoring via MyST or an expert human tutor following classroom instruction on the science topic, representing over 4.5 hr of tutoring across the 16 sessions. A quasi-experimental design was used to compare average learning gain for 3 groups: human tutoring, virtual tutoring, and no tutoring. Learning gain was measured using standardized assessments given to students in each condition before and after each science module. Results showed that students in both the human and virtual tutoring groups had significant learning gains relative to students in the control classrooms and that there were no significant differences in learning gains between students in the human and MyST human tutoring conditions. Both teachers and students gave high-positive survey ratings to MyST.
We investigate a set of techniques for RNN Transducers (RNN-Ts) that were instrumental in lowering the word error rate on three different tasks (Switchboard 300 hours, conversational Spanish 780 hours and conversational Italian 900 hours). The techniques pertain to architectural changes, speaker adaptation, language model fusion, model combination and general training recipe. First, we introduce a novel multiplicative integration of the encoder and prediction network vectors in the joint network (as opposed to additive). Second, we discuss the applicability of i-vector speaker adaptation to RNN-Ts in conjunction with data perturbation. Third, we explore the effectiveness of the recently proposed density ratio language model fusion for these tasks. Last but not least, we describe the other components of our training recipe and their effect on recognition performance. We report a 5.9% and 12.5% word error rate on the Switchboard and CallHome test sets of the NIST Hub5 2000 evaluation and a 12.7% WER on the Mozilla CommonVoice Italian test set.
We investigate a set of techniques for RNN Transducers (RNN-Ts) that were instrumental in lowering the word error rate on three different tasks (Switchboard 300 hours, conversational Spanish 780 hours and conversational Italian 900 hours). The techniques pertain to architectural changes, speaker adaptation, language model fusion, model combination and general training recipe. First, we introduce a novel multiplicative integration of the encoder and prediction network vectors in the joint network (as opposed to additive). Second, we discuss the applicability of i-vector speaker adaptation to RNN-Ts in conjunction with data perturbation. Third, we explore the effectiveness of the recently proposed density ratio language model fusion for these tasks. Last but not least, we describe the other components of our training recipe and their effect on recognition performance. We report a 5.9% and 12.5% word error rate on the Switchboard and CallHome test sets of the NIST Hub5 2000 evaluation and a 12.7% WER on the Mozilla CommonVoice Italian test set.
This article describes a comprehensive approach to fully automated assessment of children's oral reading fluency (ORF), one of the most informative and frequently administered measures of children's reading ability. Speech recognition and machine learning techniques are described that model the 3 components of oral reading fluency: word accuracy, reading rate, and expressiveness. These techniques are integrated into a computer program that produces estimates of these components during a child's 1-min reading of a grade-level text. The ability of the program to produce accurate assessments was evaluated on a corpus of 783 one-min recordings of 313 students reading grade-leveled passages without assistance. Established standardized metrics of accuracy and rate (words correct per minute [WCPM]) and expressiveness (National Assessment of Educational Progress Expressiveness scale) were used to compare ORF estimates produced by expert human scorers and automatically generated ratings. Experimental results showed that the proposed techniques produced WCPM scores that were within 3-4 words of human scorers across students in different grade levels and schools. The results also showed that computergenerated ratings of expressive reading agreed with human raters better than the human raters agreed with each other. The results of the study indicate that computer-generated ORF assessments produce an accurate multidimensional estimate of children's oral reading ability that approaches agreement among human scorers. The implications of these results for future research and near term benefits to teachers and students are discussed.Reading assessments provide school districts and teachers with critical and timely information for identifying students who need immediate help; for making decisions about reading instmction; for monitoring individual student's progress in response to instmc-
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.