Wednesday, April 01, 2020.

PLC: Packet Loss Compensation.

If audio comes in from the network on time the packet is decoded and produces a block of audio.

If a packet does not arrive in time or is dropped then the WaveNetEQ network has to predict what it would have been based on previous audio context.

If multiple packets drop then the last synthesized output is used to predict the next packet.

Trained with Teacher Forcing: during training the correct answer from ground truth is pulled instead of a prior prediction--makes it easier to train the network when it is too nubile to make its own useful prior predictions.

Network layout

Supposedly its a spectrogram -> autoregression -> spectrogram -> inverter network.