Performance analysis of distributed speech recognition over 802.11 wireless networks on the TIMIT database

This paper presents a study of the performance of an AURORA-based distributed speech recognition system over IEEE 802.11 wireless networks. The recognition features are vector quantized, packetized and sent over a simulated 802.11b network with interfering traffic. At the receiver, error concealment, VQ decoding and recognition are performed. We investigate the influence of the feature encoder, of the packetizer and of packet losses on a speaker-independent phoneme recognizer for continuous speech. The recognition system, based on hidden Markov model, is tested using a number of Gaussian mixture densities ranging from 1 to 32. We adopt the TIMIT database, ideal for evaluating phone recognizers. Network scenarios at various degrees of congestion were considered, as well as different settings of the MAC-layer standard error recovery scheme. The resulting matrix of phone recognition rates describes the performance of a 802.11 DSR system for a wide range of conditions, contributing to the understanding of DSR over 802.11 as well as to the design of advanced wireless DSR techniques.