Packet-layer Speech Quality Assessment For The AMR Networked Voice

With the popularity of wireless broadband and the development of the network，voice applications have gradually penetrated into our daily life. To meet the customers’requirements for voice quality, it is necessary to monitor the quality of networkedspeech streams in real-time at network nodes. Due to the limited performance of thenetwork nodes and the large amount of data traffics, a low complexity algorithm isneeded. The packet-layer model only uses the packet-header information withoutresorting to the related payload information, and it has low computational complexity.Therefore, the packet-layer model is well suited for real-time monitoring at networksnodes.Accordingly, no-reference speech quality assessment based on the packet-layermodel is addressed in this paper. The framework of speech quality assessment forAMR-WB+codec is developed. Firstly, by analyzing the packet header, informationabout coding rate of the bit stream, the number of lost packets and packet loss locationsis extracted to predict the voice quality. Then, coding distortion corresponding to the bitrate is determined through experiments. This paper mainly researches the effect ofpacket loss on speech quality. For different methods of dealing with packet loss, fourmethods for speech quality prediction based on the packet-layer model are established.Firstly, packet loss rate is used to measure the effect of packet loss on speech quality.The relationship among packet loss rate, coding distortion and speech quality isanalyzed and the speech quality evaluation model is proposed based on packet loss rate.Since the loss of mute frames has little impact on speech quality, the fact is wellincorporated in the proposed model. Considering the different influence of differentpacket loss patterns on the quality of speech, the packet loss burstiness parameter isobtained by analyzing the relations between three parameters, i.e., the density of burst,the number of burst periods and the ratio of burst loss within loss, and packet lossburstiness level, the speech quality evaluation model taking packet loss burstiness intoconsideration is proposed. Finally, based on three models mentioned above andaccording to the relationship between per frame and speech sequences, the two-stageMinkowski summation is used to reflect the effect of packet loss locations on speechquality, and the speech quality evaluation model based on temporal pooling is built. Theexperimental results show that the proposed four models have good performance, and the forth one using temporal pooling provides the best consistency between subjectiveand object values.