CSR: A Cloud-Assisted Speech Recognition Service for Personal Mobile Device

Abstract

Automatic speech recognition (ASR) is a technology which converts the phrases or words spoken by human into text. As a mature technology, ASR has become an alternative input method on many mobile devices, complementing the other input methods operated by hands. Although the technology has been developed for years, the accuracy and computational complexity of ASR have prohibited ASR from being used as the primary input method on mobile devices. While speaker-dependent ASR (SD-ASR) technologies may be used to improve the recognition accuracy, the user is often reluctant to take a time-consuming training process needed to enable SD-ASR for each device. To overcome these problems, we propose a cloud-assisted speech recognition service and its infrastructural design, called CSR, which utilizes servers in the cloud to accelerate ASR, integrates SD-ASR technologies to improve the accuracy of ASR, and populate SD information to enable SD-ASR on multiple mobile device. We have built a prototype to observe the benefits of the CSR service and the issues in load balance, power consumption, and privacy on the client and the server. We show that the CSR service offers fast responses, good accuracy, high availability and good scalability in serving many concurrent users.