PLCLM – Demo Page

Abstract

Packet Loss Concealment (PLC) is a fundamental component of real-time audio communication systems running over packet-switched networks. While most existing PLC approaches operate in the waveform domain, modern audio transmission pipelines increasingly rely on neural audio codecs that represent signals as sequences of discrete latent codewords. In this work, we investigate PLC directly in the compressed domain by predicting missing codewords rather than reconstructing corrupted waveforms. Leveraging the structured and low-dimensional nature of neural codec representations, we formulate PLC as a causal sequence modeling problem and propose an autoregressive Transformer trained to perform next-token prediction on codec codeword streams. The model operates strictly causally and is compatible with low-latency, real-time constraints. Using the Descript Audio Codec (DAC) as a representative neural codec, we analyze the perceptual relevance of parallel codebooks and show that accurate reconstruction can be achieved by predicting only a subset of codewords. Experimental results demonstrate that codeword-level PLC yields perceptually plausible and semantically consistent reconstructions under a wide range of packet loss rates, while providing a principled and deployment-oriented alternative to waveform-domain concealment methods.

Below is a curated set of audio samples affected by simulated packet loss ranging from 10% to 60%. The degradation is applied in the compressed domain, where missing packets are replaced with a silent prototype packet, effectively simulating zero-filling as a naive reconstruction strategy. All samples are drawn from the test set. Headphone use is recommended for optimal listening.

Clean Corrupted PLCLM restored

10% PLR