Performance Analysis of Distributed Speech Recognition Services over Noisy 802.11b Wireless Networks

The performance of an AURORA-like distributed speech recognition system over IEEE 802.11 WLANs is studied. The recognition features are packetized and sent over an 802.11b network. At the receiver recognition is performed. Two different scenarios are simulated to analyze DSR performance in presence of losses due to either low received power or to network congestion. Varying recognizer complexities, packet lengths, number of concurrent flows, and signal power levels are considered in both scenarios. Experimental results on a connected digits task show that for low signal power levels, the best recognition performance is obtained when speech features are sent in small IP packets, while in the case of network congestion the best performance is obtained by increasing the packet size.