A 2D complex convolutional layer module. This module performs a 2D convolution on complex-valued inputs by applying separate convolutions to

Short-Time Fourier Transform (STFT) module. References: - https://github.com/gemelo-ai/vocos - https://github.com/echocatzh/

A layer normalization module for 2D complex tensors.

PyTorch compatible dataset for paired audio and acoustic features.

A 1D complex convolutional layer module. This module performs a 1D convolution on complex-valued inputs by applying separate convolutions to

A 2D residual block module based on ConvNeXt architecture. Reference: - https://github.com/facebookresearch/ConvNeXt

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

A module to compute a mel-spectrogram from waveforms.

Customized collater for Pytorch DataLoader in training.

Applies an activation function to both the real and imaginary parts of a complex tensor. This class is a wrapper around a given activation f

A batch normalization module for 2D complex-valued tensors. This module applies batch normalization to both real and imaginary parts of comp

Complex-valued 2D residual block based on ConvNeXt architecture.

PyTorch compatible dataset for acoustic features.

HiFi-GAN's multi-period discriminator module. This module contains multiple PeriodDiscriminators, each operating on a different period, to c

UnivNet's multi-resolution spectral discriminator module. This module contains multiple spectral discriminators, each analyzing the input wa

HiFiGAN's period discriminator module. This discriminator operates over periodic patterns in the waveform by splitting the input into ch

UnivNet's spectral discriminator module. This module extracts features from the input waveform by converting it into the frequency domain vi

Customized trainer module for GAN-based vocoder training.

Pitch-dependent dilated 1d convolutional neural network module. This module performs 1D convolution with pitch-dependent dilation, adjusting

Pitch-dependent dilated 2D convolutional neural network module. This module performs 2D convolution with pitch-dependent dilation, adjusting

Module for calculating adversarial loss in GANs.

A batch normalization module for 1D complex-valued tensors. This module applies batch normalization to both real and imaginary parts of comp

A layer normalization module for 1D complex tensors.

A normalization module for complex tensors. This module normalizes complex-valued inputs by computing the covariance matrix between the real

Complex-valued Wavehax generator module. This class examines whether incorporating the algebraic structure of complex numbers enhances perfo

Discrete Wavelet Transform (DWT) using Haar (Daubechies 1) wavelet for 1D signals. This class performs a single-level decomposition and reco

Module for feature matching loss in GANs, comparing latent features.

Module for calculating L1 loss on Mel-spectrograms.

UnivNet's combined discriminator module. This module combines the multi-resolution spectral discriminator and the multi-period discriminator

Multi-scale Wavehax generator module. The prior waveform is first decomposed into multiple scales, transformed into time-frequency representation

A module for subscale analysis and synthesis with learnable convolutions. Reference: - Multi-stream HiFi-GAN with data-driven wavefo

Pseudo Quadrature Mirror Filter (PQMF) module. This implementation is based on the PQMF implementation used in the ParallelWaveGAN repository, wh

Auto-switching PCHIP upsampler: - If B == 1 -> uses the single-sample fast path (minimal overhead). - If B > 1 -> uses a batched/

Wavehax generator module. This module produces time-domain waveforms through complex spectrogram estimation based on the integration of