Distributed speech recognition over IP networks on the Aurora 3 database

In this paper we present the performance obtained by a Distributed Speech Recognition System (DSR) operating over simulated Internet networks. The front-end and recognizer are the standards proposed by ETSI STQ-AURORA Project; and the Aurora 3 databases were considered for training and testing. One characteristic of the IP network is the loss of packets. Here the packet transmission over IP networks was modeled by (1) random losses, (2) losses generated by a 2-state Gilbert model and (3) Bottleneck network simulations. The recognition results show that solitary losses and short burst losses do not severely affect the recognition performance. However, strongly bursty packet losses, as those generated by real Web traffic over a Internet network, can have a harmful impact on recognition performance. Therefore, so that DSR over Internet would be successful it is necessary high levels of Quality of Service.