Tencent Cloud H5 Voice Communication QoE Optimization: WebRTC QoS Techniques and Congestion Control

On April 21, Tencent Cloud+ Community held a technical salon titled “‘Sound’ comes to you, ‘Vision’ is visible – Audio‑Video Technology Development Practice”. Senior Engineer Zhang Ke from Tencent Audio‑Video Lab presented the Tencent Cloud H5 Voice Communication QoE optimization, covering the H5 solution, overall QoS optimization framework, specific techniques, and operational methods.

The QoS optimization includes four areas: bandwidth estimation and congestion control, packet‑loss mitigation, latency reduction, and jitter mitigation. Zhang highlighted interoperability issues among WebRTC‑WebRTC, TBS‑WebRTC, and TBS‑native, and introduced a retrospective analysis tool that speeds up problem discovery and iteration.

Recent standards updates were mentioned: W3C released the WebRTC standard in November, and the international group RETF published RFC 8298 in December. The speaker shared statistics from the 2017 CallStatus.io WebRTC Global Quality Report, noting that about 10% of calls drop due to bandwidth, loss, or flow issues, 10‑15% of users are dissatisfied, and roughly 7% experience severe packet loss.

Three main causes of poor quality were identified: (1) P2P connections lack guaranteed link quality across carriers; (2) poor browser interoperability and compatibility; (3) the need for a comprehensive solution, which Tencent has developed.

Two core technical advantages were described: a real‑time audio‑video solution and WebRTC support on the QQ Browser’s TBS kernel, allowing extensive code modifications and optimizations. The platform also offers streaming, recording, and on‑demand capabilities.

The system deploys over 60 global nodes in more than 30 countries, achieving sub‑36 ms latency for 98% of domestic node pairs and sub‑100 ms for 90% of global pairs. Continuous monitoring guides node placement based on user traffic.

QoS control is implemented via SFU and MCU architectures with a three‑level strategy: upstream bandwidth estimation and loss‑mitigation, downstream bandwidth estimation and loss‑mitigation, and a central decision engine that adjusts traffic based on real‑time quality metrics.

Key QoS technologies discussed include bandwidth estimation, loss‑resilient algorithms (ARQ, FEC, PLC), and adaptive congestion control. Various congestion control algorithms were compared, such as TFRC, LEBDBT, GCC (delay‑based receiver, loss‑based sender), SCReam, NADA, FRACTal, and QUIC, highlighting their strengths and weaknesses in different network conditions.

FEC mechanisms were examined in depth: in‑band FEC, XOR‑based FEC, Reed‑Solomon, interleaved coding, and Fountain codes. The trade‑offs between redundancy, latency, and burst‑loss resilience were explained, along with design considerations for dynamic redundancy rates and feedback handling.

NACK algorithm key points were listed, emphasizing jitter‑buffer‑aware requests, end‑to‑end delay control, fine‑grained retransmission evaluation, and operational monitoring.

Operational methods to reduce traffic overhead were presented: minimizing transport headers, increasing packetization intervals, lowering codec bitrate via VAD/DTX, and reducing redundancy.

Latency sources were broken down into network delay, device delay, playback delay, and codec/processing delay. The speaker stressed the importance of monitoring and optimizing each component.

Finally, the talk covered QoE evaluation, AB‑testing of new FEC strategies, and the continuous improvement cycle that has led to over a thousand bug fixes from WebRTC M56 to M66.