Rethinking Speech Recognition on Mobile Devices

In this paper, we describe our experiences and thoughts on building speech applications on mobile devices for developing countries. We describe three models of use for automatic speech recognition (ASR) systems on mobile devices that are currently used – embedded speech recognition, speech recognition in the cloud, and distributed speech recognition; evaluate their advantages and disadvantages; and finally propose a fourth model of use that we call Shared Speech Recognition with User-Based Adaptation. This proposed model exploits the advantages in all the three current models, while mitigating the challenges that make any of the current models less feasible, such as unreliable cellular connections or low processing power on mobile devices, which are typical needs of speech application in developing regions. We also propose open questions for future research to further evaluate our proposed model of use. Finally, we demonstrate the performance of two mobile speech recognizers that are either used in a lab setting to compare the recognition accuracy against a desktop, or used in real-world speech applications for mobile devices in the developing world.