Measuring the Impact of Network Performance on Cloud-Based Speech Recognition An Empirical Study of Apple Siri and Google Speech Recognition

Cloud-based speech recognition systems enhance Web surfing, transportation, health care, etc. For example, using voice commands helps drivers search the Internet without affecting traffic safety risks. User frustration with network traffic problems can affect the usability of these applications. The performance of these type of applications should be robust in difficult network conditions. We evaluate the performance of several client-server speech recognition applications, under various network conditions. We measure transcription delay and accuracy of each application under different packet loss and jitter values. Results of our study show that performance of client-server speech recognition systems is affected by jitter and packet loss, which commonly occur in WiFi and cellular networks. An experimental study on client-server speech recognition applications is reported in Impact of the network performance on cloud-based speech recognition systems, in which a solution that uses network coding to improve the performance of cloud-based speech recognition applications has been proposed. The aforementioned paper is published in ICCCN 2015 [8]. Designing and implementing of experimental testbeds by using TCP and UDP connections and also designing and implementing another testbed that uses fountain codes on UDP connection has been introduced in the paper. In this paper, we design and implement an extensive experimental evaluation of five client-server speech recognition applications to compare the performance of these applications under different network conditions.