Performance of The ETSI Distributed Speech Recognition Algorithm Over GSM and IP Networks Effekten av ETSI Algoritme for Distribuert Tale Gjenkjenning over GSM og IP nettverk

ABSTRACT: The purpose of Distributed Speech Recognition (DSR) is to improve the performance of speech recognition services across communication networks. The DSR method is to extract feature vectors at a front-end terminal and transmit them as data with error protection to a back-end server. The ETSI STQ-AURORA DSR working group has standardised such a method. In this paper the ETSI DSR algorithm is tested over GSM and IP networks with varying bit error rates and packet loss conditions. A GSM channel with poor conditions (a bit error rate of 6.7e-3) is found to reduce the recognition performance by 6.3% compared to an errorfree signal. An IP network with poor conditions (15% packet loss) reduces the recognition performance by as little as 0.9%. The test results show that the ETSI DSR algorithm achieve good recognition performance over GSM and IP networks.