A neural network based model for VoIP speech quality prediction

Abstract

It is increasingly important to model the VoIP speech quality. Network factors(e.g packet loss) and source impairments (e.g. codec type) should be considered in any proposed solution. Some new factors's affection, such as jitter standard deviation, is recently studied. In this paper, we proposes a new neural network models for predicting VoIP speech quality. The proposed approach use intrusive methods (PESQ) for neural network training, which avoids time-consuming subjective tests.Our method aims to overcome the limitations of the available neural model in the literature, and it presents several advantages over them: new network parameters (jitter standard deviation ) and new source parameters (language) are considered in our approach; different network simulation system is setup . We used latest NISTnet, which is believed more realistic for modeling actual network than Gilbert model or other simulation system . The model experiment results suggested that the designed neural network model works well for speech quality.