This paper analyzes the impact of signal phase handling in one of the most popular architectures for the generative synthesis of audio effects: variational autoencoders (VAEs). Until quite recently, autoencoders based on the Fast Fourier Transform routinely avoided the phase of the signal. They store the phase information and retrieve it at the output or rely on signal phase regenerators such as Griffin--Lim. We evaluate different VAE networks capable of generating a latent space with intrinsic information from signal amplitude and phase. The Modulated Complex Lapped Transform (MCLT) has been evaluated as an alternative to the Short-Time Fourier Transform (STFT). A novel database on beats has been designed for testing the architectures. Results were objectively assessed (reconstruction errors and objective metrics approximating opinion scores) with autoencoders on STFT and MCLT representations, using Griffin--Lim phase regeneration, multichannel networks, as well as the Complex VAE. The autoencoders successfully learned to represent the phase information and handle it in a holistic approach. State-of-the-art quality standards were reached for audio effects. The autoencoders show a remarkable ability to generalize and deliver new sounds, while overall quality depends on the reconstruction of phase and amplitude.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21885
(513KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
