A Causal Convolutional Approach for Packet Loss Concealment in Low Powered Devices

This paper presents a deep learning model for audio Packet Loss Concealment (PLC) for real time communications that is accurate, lightweight, with a low inference time suitable for low powered mobile handsets. We leverage dilated causal convolutions to track short term time dependent features of previous audio making the architecture fully convolutional. The model is semi-autoregressive, meaning it can work autoregressively and non-autoregressively depending on audio loss length and model output size. Whilst existing solutions can perform PLC up to 120 ms. our proposed model can perform PLC for packet losses up to 200ms with an inference time of 51ms on a CPU and a model size of 4.19 MB in Tensorflow Lite. We also show how the inference time can be decreased by increasing the model output size without any decrease in model accuracy or significant increases in model size. The model is assessed in terms of RMSE, PLC-MOS, STOI, PESQ and inference time and compared to two baseline methods.