Combinational methods for improvement of packet-loss recovery in distributed speech recognition

In distributed speech recognition, speech features are transmitted inside packets from a channel to speech recognition server. Some problems such as channel noise, packet loss and delay decrease distributed speech recognition rate. Several methods including packet-loss recovery approaches have been proposed to overcome to this decrease in recognition rate. Most of packet-loss recovery methods, don´t notice to type of information existing in packets. In this paper, we propose two methods for packet-loss recovery based on type of existed information (speech features) in packets. In addition, we combine these two methods with conventional packet-loss recovery methods such as repetition, interleave and interpolation. We evaluate the combinational methods in two cases: random packet loss and packet loss based on Gilbert model. Recognition results on isolated words of TIDIGITS show that proposed combinational methods outperform the conventional methods.