Evaluating on-device ASR on Field Recordings from an Interactive Reading Companion

Abstract

Many applications designed to assess and improve oral reading fluency use automated speech recognition (ASR) to provide feedback to students, teachers, and parents. Most such applications rely on a distributed architecture with the speech recognition component located in the cloud. For interactive applications, this approach requires a reliable Internet connection that may not always be available. We investigate whether on-device ASR can be used for a virtual reading companion using recordings obtained from children both in a controlled environment and in the field. Our limited evaluation makes us cautiously optimistic about the feasibility of using on-device ASR for our application.