A WLAN Scheduling Algorithm to Reduce the Energy Consumption of a Distributed Speech Recognition Front-End

In distributed speech recognition, speech features are computed on a mobile device, compressed, and sent to a server that performs the computationally intensive search for the most likely word sequence. Much of the current research in distributed speech recognition has been in the area of feature compression and communication robustness over wireless links, including error correction and concealment techniques. However, another challenge in designing a distributed speech recognition system is minimizing the energy consumption on the mobile device. We consider quality-of-service tradeoffs including compression ratio and overall system latency. Our measurements verify that for high speed wireless interfaces such as 802.11b, small changes in compression rates have little effect on system level energy consumption. However, for wireless networks with lower power/bit-rate ratios such as Bluetooth, the choice of bit-rate and compression ratio becomes more important. We present a wireless LAN scheduling algorithm to minimize the energy consumption of a distributed speech recognition front-end on a mobile device. By powering down the 802.11b interface when not in use, we are able to reduce the energy consumption by up to a factor of 5 in heavy traffic conditions. Increasing the total amount of time spent in the off state to almost one second will allow the system to save power regardless of traffic conditions. We compare the results of this power saving algorithm to the low-power mechanisms of Bluetooth. The lower overhead of Bluetooth allows for greater energy savings with a much lower delay of approximately 300ms.