Quality-based playout buffering with FEC for conversational voIP

Qipeng Gong, Peter KabalInterspeech 2010

TL;DR

This paper tackles the challenge of maintaining high voice quality in VoIP by proposing a novel quality-based playout buffering algorithm that integrates Forward Error Correction (FEC). By intelligently adapting to network conditions and using FEC to proactively correct for packet losses, the algorithm significantly improves perceived conversational quality and interactivity without increasing delay.

What Problem Does This Paper Address?

Voice over IP (VoIP) offers a flexible way to transmit voice, but it faces significant challenges compared to traditional phone lines. The "best-effort" nature of IP networks means that packets can be delayed (jitter) or lost entirely, leading to choppy audio and dropped words. While jitter buffers are used to smooth out delays and packet loss concealment (PLC) techniques try to fill in the gaps, these methods often come at the cost of increased conversational delay, which harms the natural flow and interactivity of a conversation. Furthermore, bursty packet losses, where multiple packets are lost in quick succession, are particularly detrimental to perceived quality, even with PLC. This paper aims to address the trade-off between minimizing delay and mitigating packet loss, especially burst losses, to achieve a better conversational experience.

Methods

The core of this paper is a two-pronged approach:

  1. Quality-Based Playout Scheduling: The authors leverage a previously developed quality measurement that considers both voice quality (using the ITU-T E-model) and conversational delay. The playout scheduling algorithm aims to optimize this combined metric. It operates on a per-talkspurt basis, meaning it adjusts buffering strategies at the beginning and end of speech segments.
  2. Beginning of Talkspurt: The algorithm uses a technique called Packet WSOLA (Wavelet Packet-based Time-Scale Modification) to stretch incoming speech packets. This effectively increases the jitter buffer size, allowing more time to receive subsequent packets and recover from potential losses. The stretch factor is adaptively controlled based on network delay information.
  3. End of Talkspurt: When a "hangover" (silence) is detected, the algorithm uses PWSOLA to compress the buffered speech, reducing the jitter buffer size and thus the overall conversational delay.
  4. Buffer-Aware Signal Processing FEC (SP-FEC): To further combat packet loss, the paper introduces a novel SP-FEC scheme. Unlike traditional FEC that might send redundant data indiscriminately, this approach is "buffer-aware."
    • Sender Side: The sender piggybacks redundant information from previous voiced packets onto the current packet. Crucially, the number of previous packets included (m) is dynamically determined based on information about the receiver's jitter buffer size, which is communicated via RTCP packets. This ensures that redundant data is only sent when it's likely to be useful for recovery. Piggybacking stops when silence (hangover) is detected.
  5. Receiver Side: The receiver sends an RTCP packet to inform the sender about its jitter buffer status if it changes significantly. This allows the sender to adjust its FEC strategy accordingly.

For simplicity in their simulations, the authors used the G.711 codec for both primary and redundant information.

Datasets

The paper uses three network delay trace files for evaluation:

  • Trace 1: Collected between McGill University (Canada) and Shanghai JiaoTong University (China) in January 2009.
  • Trace 2 & 3: Obtained from a previous study, connecting the UK and China.

The network packet loss is modeled using a 2-state Gilbert Model with a 2% network loss rate and an expected burst loss of 2 packets. The conversational audio used in simulations is from a real dialog with noisy background, featuring an "ask-response" pattern.

Key Insights and Ideas

  • Burstiness is a Killer: The paper empirically demonstrates that longer burst lengths of packet loss significantly degrade perceived voice quality (PESQ-MOS scores), even with standard PLC techniques. This highlights the need for methods specifically designed to handle bursty losses.
  • FEC Needs to be Smart: Simply adding FEC can increase overhead and delay. The key innovation here is making FEC "buffer-aware." By coordinating the amount of redundant information sent with the receiver's buffer status, the system avoids wasting bandwidth and processing on data that won't be usable.
  • Delay and Quality are Intertwined: The paper reinforces the understanding that for conversational VoIP, both low delay and high voice quality are critical. The proposed algorithm explicitly optimizes for both, recognizing that a long delay can make even perfect audio sound unnatural.
  • Adaptive Buffering is Key: The dynamic adjustment of the jitter buffer size at both the beginning and end of talkspurts, combined with the adaptive FEC, allows the system to be responsive to changing network conditions without sacrificing interactivity.

Key Results

The proposed algorithm, termed VJM_FEC, was compared against several other jitter buffering algorithms, including Exp-Avg, Fast-Exp, Sun_opt, and their previous work, VJM_adaptive.

  • Improved Perceived Quality: VJM_FEC consistently achieved higher PESQ-MOS scores than VJM_adaptive, indicating better perceived voice quality.
  • Comparable Conversational Delay: Importantly, VJM_FEC achieved these quality improvements with similar conversational delays to VJM_adaptive, demonstrating that the FEC integration did not introduce significant new latency.
  • Reduced Burst Losses: Table 3 shows that VJM_FEC was more effective at reducing the number of burst loss packets compared to VJM_adaptive, with reductions ranging from 7.64% to 13.39% across the tested traces.
  • Balancing Act: The results suggest that quality-based algorithms (like VJM_FEC and VJM_adaptive) generally achieve a better balance between packet losses and delays compared to simpler averaging methods (Exp-Avg, Fast-Exp), which often sacrifice one for the other.

Limitations and Future Work

  • Codec Simplification: The paper used G.711 for both primary and redundant information for simplicity. Future work could explore the benefits of using different codecs for the redundant stream (e.g., a lower-bitrate codec) to further reduce overhead.
  • FEC Scheme Details: While the buffer-aware SP-FEC is a key contribution, the paper mentions that the exact implementation of compact redundant encoding could involve a small quality loss during recovery. Further research could optimize this trade-off.
  • Real-World Deployment: The results are based on simulations using recorded delay traces. While robust, real-world testing on live networks would be the ultimate validation.
  • VAD/DTX Integration: The paper mentions using VAD/DTX for hangover detection. Further exploration into the nuances of these detection mechanisms and their impact on FEC and playout scheduling could be beneficial.

Summary updated

BibTeX
@inproceedings{gong2010quality,
  author = {Qipeng Gong and Peter Kabal},
  title = {{Quality-based playout buffering with FEC for conversational voIP}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure