A learned compression model that encodes audio into discrete or continuous latent representations optimized for reconstruction.