FastMVAE: A Fast Optimization Algorithm for the Multichannel Variational Autoencoder Method

FastMVAE2: On improving and accelerating the fast variational autoencoder-based source separation algorithm for determined mixtures

“…While the FastMVAE method can significantly reduce the computation time compared to the MVAE method, its source separation accuracy has been confirmed to be somewhat less than that of the MVAE method [37]. We believe that this is for j = 1 to J do 4:…”

“…We chose ILRMA [7], the MVAE method [19] 1 , and the FastMVAE method [37] as the baseline methods for both the speaker-dependent and speaker-independent cases, and IDLMA [23] as another baseline method only for the speaker-dependent scenario. For all the methods, the parameter optimization algorithms were run for 60 iterations, and the separation matrix W(f ) was initialized with an identity matrix.…”

“…Note that unlike other methods where speaker information is estimated, IDLMA requires speaker information in order to properly select the corresponding pre-trained network. The network architectures for the CVAE and ACVAE source models were the same as those used in [37], where the encoder consisted of 2 convolutional layers using GLU following a regular convolutional layer, the decoder consisted of 2 deconvolutional layers using GLU following a regular deconvolutional layer, and the classifier consisted of 3 convolutional layers using GLU following a regular convolutional layer. All the GLU layers used batch normalization to stabilize the training.…”

“…To address this problem, we previously proposed a fast algorithm called "FastMVAE" [37], which uses an auxiliary classifier VAE (ACVAE) [36] to model the generative distribution of source spectrograms. In this method, the encoder and auxiliary classifier are trained in such a way that they learn to infer the latent space variables and class variables, respectively, given a spectrogram.…”