How Context‑Based PLC Boosts VoIP Quality in Weak Networks

Why VoIP Calls Stutter

In IP‑based voice calls, weak network conditions often cause packet loss, leading to 40‑100 ms interruptions that can create confusing misunderstandings.

Common Remedies: FEC and PLC

Traditional solutions include Forward Error Correction (FEC), which adds redundant data to each packet so lost packets can be reconstructed, but it consumes extra bandwidth and may face encoder compatibility issues.

Packet Loss Concealment (PLC) operates on the decoder side, predicting missing frames from previously received audio without additional bandwidth. Classic PLC, such as the OPUS‑PLC used in many codecs, can only restore a small fraction of a Chinese character’s duration, resulting in limited perceptual improvement.

Deep‑Learning‑Based PLC

Recent attempts use deep learning—spectral regression or generative models—to predict missing spectra or time‑domain samples, allowing compensation for up to 120 ms of continuous loss. However, these models often exceed 10 MB and demand high CPU usage, making them unsuitable for low‑end smartphones.

Tencent’s Context‑Based PLC (cPLC)

cPLC leverages the highly structured nature of speech, modeling short‑term contextual features to keep the decoder model under 600 KB. Tests show negligible CPU overhead even on mid‑range devices, meeting the low‑latency and low‑complexity requirements of real‑time audio applications.

Evaluation used a third‑party network impairment device to simulate various loss patterns and the ITU‑T P.863 MOS metric (range 1‑4.75). Across all scenarios, cPLC achieved MOS improvements of 0.1‑0.2 over baseline and consistently outperformed OPUS‑PLC, especially in burst‑loss cases.

Results

In weak‑network or burst‑loss environments, cPLC provides more reliable voice quality, a benefit confirmed after its deployment in Tencent Meeting.

Conclusion

cPLC is Tencent Meeting’s next‑generation packet loss compensation solution, blending classic signal‑processing with lightweight deep‑learning to deliver continuous‑loss concealment while keeping computational cost low, thereby enhancing user experience in challenging network conditions.