Utilizing the neural networks for speech quality estimation based on the network characteristics

Abstract

The paper deals with an issue of the speech quality estimation in Voice over IP technology under packet loss. Packet loss is a major problem for real-time Internet applications, we applied four-state Markov model for modeling the impact of network impairments on speech quality, afterwards, the resilient back propagation (Rprop) algorithm was used to train a neural network. The general and RFC3611-compliant solution, which allows for quick and precise speech quality estimation without the need to analyze or model the voice signal carried by the RTP (Real-time Transport Protocol) packets, is the contribution of this paper. The proposed solution is tested on G.711 A-law and further generalizes the already presented concepts of the speech quality estimation in the IP environment. The proposed approach of speech quality assessment belongs to non-intrusive methods and is based on the back-propagation neural networks.