Performance improvement of distributed Turkish continuous speech recognition system in case of burst packet losses

In this paper, effect of burst packet losses on performance of continuous speech recognition over IP network is investitaged. Here distributed Turksih continuous speech recognition system is considered and burst packet losses are artificially generated. In the system, speech feature vectors are transmitted over UDP with a special header. Two packet loss concealment methods are employed: Lagrange and spline interpolation. In the experimental study, the speech feature vectors are extracted using Hidden Markov Tool Kit (HTK), and SRI Language Modelling toolkit is used to generate statistical language models. The acoustic modeling and recognition are performed using AT&T software. The experimental results show that the word error rate (WER) obtained by spline interpolation for the burst packet losses is better than that of Lagrange interpolation method. The WER for Lagrange method increases faster than spline method's WER when the burst duration is increased.