speechbrain.lobes.models.segan_model module

This file contains two PyTorch modules which together consist of the SEGAN model architecture (based on the paper: Pascual et al. https://arxiv.org/pdf/1703.09452.pdf). Modification of the initialization parameters allows the change of the model described in the class project, such as turning the generator to a VAE, or removing the latent variable concatenation.

Loss functions for training SEGAN are also defined in this file.

Authors

• Francis Carter 2021

Summary

Classes:

Discriminator	CNN discriminator of SEGAN
Generator	CNN Autoencoder model to clean speech signals.

Functions:

d1_loss	Calculates the loss of the discriminator when the inputs are clean
d2_loss	Calculates the loss of the discriminator when the inputs are not clean
g3_loss	Calculates the loss of the generator given the discriminator outputs

Reference

class speechbrain.lobes.models.segan_model.Generator(kernel_size, latent_vae, z_prob)

Bases: Module

CNN Autoencoder model to clean speech signals.

Parameters:

• kernel_size (int) – Size of the convolutional kernel.

• latent_vae (bool) – Whether or not to convert the autoencoder to a vae

• z_prob (bool) – Whether to remove the latent variable concatenation. Is only applicable if latent_vae is False

forward(x)

Forward pass through autoencoder

class speechbrain.lobes.models.segan_model.Discriminator(kernel_size)

Bases: Module

CNN discriminator of SEGAN

Parameters:

kernel_size (int) – Size of the convolutional kernel.

forward(x)

forward pass through the discriminator

speechbrain.lobes.models.segan_model.d1_loss(d_outputs, reduction='mean')

Calculates the loss of the discriminator when the inputs are clean

speechbrain.lobes.models.segan_model.d2_loss(d_outputs, reduction='mean')

Calculates the loss of the discriminator when the inputs are not clean

speechbrain.lobes.models.segan_model.g3_loss(d_outputs, predictions, targets, length, l1LossCoeff, klLossCoeff, z_mean=None, z_logvar=None, reduction='mean')

Calculates the loss of the generator given the discriminator outputs