Exploring Representation Autoencoders for Text-to-Audio Generation

Abstract

Representation autoencoders (RAE), created by pairing a frozen pretrained self-supervised learning (SSL) encoder with a trained decoder, have recently demonstrated state-of-the-art performance on image generation. Their use in the audio domain, however, remains limited: text-to-audio (TTA) models are typically trained in the latent space of dedicated autoencoders where the encoder and the decoder are trained jointly with reconstruction objectives. This paper presents a study on audio representation autoencoders (AuRA) and their applicability to general audio TTA. We train decoders for two pretrained SSL audio encoders and evaluate their reconstruction quality against autoencoders used by recent TTA systems. After obtaining decoders for the SSL representations, we also train rectified flow models in the SSL and baseline latent spaces to isolate the autoencoder's effect on generation. We find that generation in SSL spaces achieves comparable or superior performance to generation in the baseline spaces, despite differences in reconstruction performance. These results highlight that reconstruction quality alone is not a sufficient proxy for generative performance.