Towards Perceptual Loss: Using a Neural Network Codec Approximation as a Custom Loss Metric

We train a neural network to approximate a low bit-rate codec and suggest the use of this network as a custom loss function. We demonstrate perceptual improvement in reconstruction quality in a secondary, autoencoding task.

Part I: Modeling a Codec as a Loss Function

	Track 1	Track 2	Track 3
Lossless
FFMPEG
Mag. STFT (FFMPEG phase)
Time Domain

Part II: Using the Loss Function with an Autoencoding Task

	Track 1	Track 2	Track 3
Lossless
Model A
Model B
Model C
Model D