Speech-Quality-Affecting Parameters

Based on the statistics collected from the above experiment, we have selected values for the network and encoding parameters as follows:

1. Loss Rate: 0, 5, 10, 20 and 40 %. In fact, the loss rate depends on the bandwidth, network load, congestion and the choice of a strict playback time. For more details about packet loss dynamics, see [103]. Of course, one can use any mechanism of FEC [18,19,116] to reduce the amount of total loss but on the cost of extra delay, which in turn controls the loss rate if strict playback time mechanism is used [37]. Other techniques for error concealment are given in Section 3.3.2. In networks employing some kind of QoS mechanisms like IP DiffServ or ATM, one can control the amount of losses to achieve certain levels of audio/speech quality [11].
2. Loss Distribution: we have chosen the number of consecutively lost packets as the loss distribution, which varies from 1 up to 5 packets dropped at a time. Of course in some situations, there may be more than 5-consecutively-lost packets. However, for the sake of simplicity, we consider them as multiple of 5-consecutively-lost packets.
3. Packetization Interval: We have chosen the values of the packetization intervals as 20, 40, 60 and 80 ms. In fact the majority of the existing real-time applications use one or more of these values. It should be mentioned that people prefer to use small values of packetization intervals to avoid the degradation of the quality due to the increase of the overall delay. On the other hand, using smaller values reduces the effective network utilization. For example, for 20-ms packetized GSM voice, the size of each packet is 33 bytes. When using the RTP/UDP/IP stack, the minimum length of the header of each packet is (12+8+20=40 bytes). In this case, the network utilization factor is bad (cf. Section 8.3.1, page ).
4. Encoding Algorithms: Concerning the speech encoding algorithms, we have selected PCM (64kbps), G726 ADPCM (32kbps) and GSM-FR (13.2kbps). For French language experiment, we did not consider the ADPCM codec. The corresponding packet sizes are 160, 80, and 33 bytes for a 20ms packetization interval. From the literature [100], the corresponding subjective MOS quality rates are 4.4, 4.1 and 3.6 respectively for English language. These values correspond to an absolute score that evaluates the codec without other impairments. These codecs are the most widely used ones over the public Internet in real-time talks and chat applications.