Music2Latent: Consistency Autoencoders for Latent Audio Compression

TL;DR

Music2Latent introduces a novel end-to-end consistency autoencoder for audio that achieves high-fidelity, single-step reconstruction and outperforms existing models in sound quality and accuracy, advancing generative audio and MIR applications.

Contribution

It presents the first successful end-to-end consistency autoencoder for audio, integrating frequency-wise self-attention and learned scaling for improved performance.

Findings

Outperforms existing autoencoders in sound quality and reconstruction accuracy

Enables high-fidelity single-step audio reconstruction

Achieves competitive results on MIR tasks

Abstract

Efficient audio representations in a compressed continuous latent space are critical for generative audio modeling and Music Information Retrieval (MIR) tasks. However, some existing audio autoencoders have limitations, such as multi-stage training procedures, slow iterative sampling, or low reconstruction quality. We introduce Music2Latent, an audio autoencoder that overcomes these limitations by leveraging consistency models. Music2Latent encodes samples into a compressed continuous latent space in a single end-to-end training process while enabling high-fidelity single-step reconstruction. Key innovations include conditioning the consistency model on upsampled encoder outputs at all levels through cross connections, using frequency-wise self-attention to capture long-range frequency dependencies, and employing frequency-wise learned scaling to handle varying value distributions across frequencies at different noise levels. We demonstrate that Music2Latent outperforms existing continuous audio autoencoders in sound quality and reconstruction accuracy while achieving competitive performance on downstream MIR tasks using its latent representations. To our knowledge, this represents the first successful attempt at training an end-to-end consistency autoencoder model.