RAVE perceptual quality must be judged by ear, since adversarial loss values do not track it

Reconstruction distance metrics track training progress but correlate poorly with perceptual audio quality, especially in phase 2 where the adversarial losses (loss_dis, loss_gen, pred_true, pred_fake) are, like most GAN losses, harder to read and do not move monotonically with quality. Because the generator and discriminator are in competition, their absolute values indicate equilibrium, not goodness. The reliable signal is the validation audio logged to TensorBoard: you listen to successive validation samples and stop when the audio sounds good, not when a number plateaus. A model producing repetitive or washed-out output signals instability that a loss curve may not reveal.

Examples

Open TensorBoard → listen to the validation audio samples across epochs and compare against reference dataset audio; do not rely on the GAN loss numbers, which are hard to read.

Assessment

Your RAVE phase 2 run shows stable loss values but the exported model sounds dull and washed-out. What monitoring approach should you have used instead, and what can you do now?