Packet delay distribution model for investigating delay of network speech recognition

Unlike multimedia streaming applications that require a smooth playback at the client, application of network speech recognition (NSR) that recognizes speech signal in a sentence-by-sentence manner might tolerate an acceptable delay. The acceptable delay is a user-defined time in which the entire sentence data should be received by the server. We proposed a calculation method to investigate the acceptable delay of network speech recognition that employs a speech segmenter to send speech signal sentence-by-sentence over TCP channel to the server. The calculation multiplies the mean packet delay of TCP flow at steady-state with the number of created packets. For validation we implemented a MATLAB program and solved it using 2500 Indonesian speech sentences. The results were then compared with the results of our previous model that used a transient analysis method. It was found that this calculation method is not appropriate due to the transient behavior of the streaming sentences.