Real-time voice communication over the internet using packet path diversity

Yi J. Liang, Eckehard Steinbach, Bernd GirodProceedings of the Ninth ACM International Conference on Multimedia - MULTIMEDIA '01 2001 · DOI

TL;DR

This paper proposes a novel approach for real-time voice communication over the internet by transmitting multiple redundant descriptions of the voice stream over different network paths. This "packet path diversity" significantly improves the trade-off between delay, packet loss, and speech quality compared to traditional single-path methods, even those using forward error correction.

What Problem Does This Paper Address?

Real-time voice communication over the internet, especially over best-effort networks like today's internet, faces a fundamental challenge: maintaining high quality despite unpredictable network conditions. These conditions include significant variations in delay (jitter) and packet loss. For voice, excessive delay means packets arrive too late to be played out, effectively becoming "late loss." Traditional methods to combat packet loss, such as Forward Error Correction (FEC) and interleaving, often increase latency or are inefficient in the face of correlated packet losses (bursts). This paper addresses the need for a method that can effectively reduce both delay and loss without sacrificing speech quality, by leveraging the inherent unreliability of the internet in a new way.

Methods

The core of the proposed solution lies in two key components:

  • Multiple Description Coding (MDC): Instead of sending a single stream, the voice signal is encoded into multiple, independent descriptions. If one description is lost, the other can still be used to reconstruct the signal, albeit with degraded quality. The paper uses a specific MDC scheme where even samples are quantized finely and the difference between adjacent even/odd samples is quantized coarsely for one stream, and vice-versa for the second stream. This provides redundancy without a significant increase in data rate.
  • Adaptive Multi-Stream Playout Scheduling: At the receiver, a novel scheduling algorithm dynamically determines when to play out packets from different streams. This algorithm uses a Lagrangian cost function to balance three competing objectives: minimizing end-to-end delay, minimizing the probability of losing both descriptions (packet erasure), and minimizing the probability of losing only one description (speech quality degradation). The algorithm learns from past packet delays to predict future loss probabilities and adjust playout deadlines accordingly.

The paper also discusses how path diversity can be achieved, either through overlay networks or future IP protocols like IPv6, and emphasizes the importance of selecting paths with uncorrelated delay and loss characteristics.

Datasets

The paper describes two main evaluation methods:

  • Internet Experiments: Two experiments were conducted over the internet.
  • Experiment 1: Source at Netergy Networks (Santa Clara, CA) to MIT (MA), using a default path and an alternative path routed through Harvard.
  • Experiment 2: Source at Rutgers University (NJ) to Erlangen University (Germany), using a default path and an alternative path routed through Harvard.
  • Data collected included packet sequence numbers and delays.
  • Network Simulations: The ns-2 network simulator was used to evaluate the scheme under various network topologies and loads.
  • Topologies: Independent paths and paths sharing a common link.
  • Traffic: Constant Bit Rate (CBR) voice streams and TCP traffic to simulate network load.
  • Parameters varied: Network load (link utilization), bandwidth of shared links, and propagation delay/standard deviation of links.

Key Insights and Ideas

  • Leveraging Unreliability: The most striking insight is that the internet's inherent unreliability (unpredictable delay and loss) can be turned into an advantage. By sending redundant information over different paths, the probability of both paths experiencing a severe disturbance simultaneously is low.
  • Uncorrelated Paths are Key: The effectiveness of path diversity hinges on the paths having largely uncorrelated delay and loss characteristics. This means that a problem on one path is unlikely to affect the other.
  • Trade-off Optimization: The Lagrangian cost function provides a principled way to balance the competing goals of low delay, low loss, and good speech quality, allowing for flexible tuning based on application requirements.
  • Delay Jitter as a Major Culprit: The experiments highlight that for real-time voice, delay jitter (and the resulting late loss) is often a more significant problem than outright packet erasure. Path diversity effectively mitigates this.
  • MDC + Path Diversity > FEC: The combination of MDC and path diversity outperforms FEC-protected single-path transmission in terms of both delay and loss reduction, without compromising speech quality.

Key Results

  • Significant Delay and Loss Reduction: Internet experiments showed substantial reductions in mean end-to-end latency and loss rates compared to FEC-protected single-path transmission. For example, at a fixed average delay of 70ms, total erasure rate was reduced from over 16% to less than 2%.
  • Improved Speech Quality (PESQ): The proposed multi-stream scheme achieved higher PESQ scores (a measure of speech quality) than single-stream FEC, indicating that the improved delay/loss trade-off did not come at the cost of degraded quality. In Experiment 1, PESQ scores were more than 0.4 higher.
  • Burst Loss Mitigation: Burst loss rate was significantly reduced, from over 3.5% to 0.5% in one experiment, which is crucial for concealing errors in voice communication.
  • Simulation Validation: Simulations confirmed these findings, showing that packet erasure rates could be kept below 4% even at high network loads (90%) with two independent paths, compared to 16% for a single stream. Burst loss was kept below 1%.
  • Shared Link Impact: While sharing a link increased delay correlation between paths, path diversity still offered significant improvements in loss reduction, especially at higher loads.

Limitations and Future Work

  • Path Correlation: The effectiveness of the scheme is highly dependent on the statistical correlation between the chosen paths. If paths become highly correlated (e.g., due to a shared bottleneck link), the benefits diminish.
  • Alternative Path Delay: If the alternative path has a significantly higher average propagation delay than the default path, the gain from path diversity becomes small.
  • Implementation Complexity: While the paper discusses overlay networks and future IP protocols, implementing true path diversity in current internet infrastructure can be challenging due to lack of explicit source routing support.
  • Future Work: The authors suggest that future work could explore more advanced MDC schemes, investigate the impact of different network topologies and shared link configurations more deeply, and further refine the adaptive playout scheduling algorithm. The potential of IPv6's source routing for simplifying path diversity implementation is also highlighted.

Summary updated

BibTeX
@inproceedings{liang2001real,
  author = {Yi J. Liang and Eckehard Steinbach and Bernd Girod},
  title = {{Real-time voice communication over the internet using packet path diversity}},
  booktitle = {Proceedings of the Ninth ACM International Conference on Multimedia - MULTIMEDIA '01},
  year = {2001},
  doi = {10.1145/500203.500205},
}
Figure