Communications Network Research Institute

EQUAL Project

Mirosław (Mirek) Narbutt Piotr Soboński

Project goal

Real-time implementation of the ITU-T E-Model quality contours in a playout buffer module of a VoIP application.

Summary

If you would like to read a short summary of the EQUAL project please click the following link: Download EQUAL summary

Project duration

October 2007 − December 2008

Funding Agency

Enterprise Ireland (under the Commercialisation Fund Proof of Concept 2007 programme). The project has been co-funded by the European Regional Development Fund (ERDF) under the Southern & Eastern Regional Operational Programme 2007-13.


Investing in your future.

Background

An extension to the ITU-T E-model developed for the pruposes of predicting VoIP transmission quality has recently been adopted by the ITU-T as Appendix I to the ITU-T G.109 Recommendation [1], [2], [3], [4]. To date, this method has been realized as an off-line application layer tool for evaluating various playout buffer algorithms [6], [7] and has been shown to be particularly effective in assessing the performance of Voice over WLAN systems [8], [9], [10], [11], [12], [13], [14]. This project proposes to develop a real-time implementation of this method that is capable of being integrated into adaptive tuning schemes for optimising VoIP quality.

Problem

A fundamental trade-off exists between average buffering delay and late packet loss. A large buffering time causes an increase in the overall delay but decreases the packet loss, while a small buffering time decreases the delay but increases the packet loss. A good playout algorithm should be able to keep the buffering as short as possible while minimizing the number of packets that arrive too late to be played out. These two conflicting goals have led to various adaptive playout algorithms, which can be grouped into four categories:

reactive algorithms that perform continuous estimation of network delays and jitter to calculate playout deadlines [14], [15], [16];
histogram-based algorithms [histogram b. alg.] that maintain a histogram of packet delays and choose the optimal playout delay from that histogram [17], [18];
algorithms that monitor packet loss ratio or buffer occupancy and adjust the playout delay accordingly [19];
algorithms that aim to maximise user satisfaction [20], [21], [22];

Although this statistical method of evaluating playout algorithms can be useful, it does not provide a direct link to perceived conversational speech quality. Instead, we predict user satisfaction from time varying loss/delay impairments using the ITU-T standardized quality contours.

Objective

The main objective of this project is to realise an adaptive de-jitter playout buffer for VoIP applications that is controlled by the real-time implementation of the E-model and which seeks to achieve the highest user satisfaction under all network conditions. The intention is to control the playout buffer module using the voice transmission quality predictor in order to minimize the effect of transmission impairments and thus maximize user satisfaction.

Implementation

Implementation has been divided on two parts:

A real-time implementation of ITU-T G.107/G.109 App. I method in RTP tools software integrated into a playout buffer module [23].
A real-time implementation of an adaptive de-jitter buffer on a selected VoIP stack.

Results

Experimental test bed

Both modified RTP tools applications (i.e. RTPsend and RTPdump) were tested through network emulations, and in a real wireless LAN environment. Realistic voice sources were used with active and silent periods (in accordance with ITU-T recommendation P.59 [5]). Voice packets were generated at the sender every 10ms during active periods. No packets were generated during silent periods. The total duration of each test was 1 hour. The simulation scenarios (see Figure 1) covered a wide range of network conditions regarding delay and jitter behaviour.

For network emulations the NISTnet 2.1.0 network emulation software has been chosen [24]. Both RTPsend and RTPdump applications were running in different networks. Voice traffic flow was sent from RTPsend to RTPdump application through a NISTnet router that modelled various delay patterns. Figure 2 shows delay patterns observed during network emulations.

Figure 1: Test simulation setup for the EQUAL module.


(A)	(B)			(C)

Figure 2: Delay patterns used in experiments: (A) delay = 0ms, jitter varies; (B) both delay and jitter vary; (C) delay = 100ms, jitter varies.

Experimental results

Both the E-Model method and the playout buffer module have been successfully implemented in the RTP tools software. We tested its performance through network simulations and via experiments in a real wireless environment.

We compared the performance of reference playout algorithms (with fixed playout deadlines, Ramjee’s [14],and Moon’s [18]) with our optimized algorithm. For comparison purposes we have chosen network traces: A, B, C. Table 1 compares the maximal values of rating factor R achieved by the reference algorithms with the values of R obtained using our optimized algorithm. As an example graphical results from fixed playout are shown below for trace A, B and C.

Algorithm	Transmission factor R
Algorithm	Trace A	Trace B	Trace C
Fixed playout delay (160 ms)	74.27	62.97	19.96
Ramjee’s alg.	69.34	79.32	63.15
Moon’s alg.	75.89	78.91	63.76
Optimized R (our solution)	89.27	87.06	75.82

Table.1. Algorithm performance comparison

Results have shown that our quality driven buffer adaptation provides the highest R score among examined playout adaptation schemes.

Key Benefits

This tool proves the concept of cross-layer design through the example of tuning the playout buffer at the application layer and will be a key step to a broader solution of multilayer tuning for wireless VoIP stacks.
This tool will allow content developers to determine the optimal parameters for tuning to ensure that highest user satisfaction under all network conditions will be achieved.

Prototype Details/Requirements

C/C++ application under Windows and Linux
VoIP stack will be used in the final implementation
Support for E-Model and quality contours standards ITU-T G.107, G.109 App. I, G.113