Initialize the ComplexConv1d module. Args: in_channels (int): Number of input channels. out_channels (int):

Reads a dataset from an HDF5 file. Args: hdf5_name (str): Path to the HDF5 file. hdf5_path (str): Dataset path within the HD

Read lines from a text file, removing newline characters. Args: file_list (str): Path to the text file containing filenames. Re

Initialize the NormLayer module. Args: channels (int): Number of input features. eps (float, optional): A sm

Initialize the PeriodDiscriminator module. Args: period (int): Period to split the waveform for 2D convolution processin

Perform the inverse Short-Time Fourier Transform (iSTFT) to reconstruct the waveform from the complex spectrogram. Args:

Normalize the real and imaginary parts of a complex tensor. Args: real (Tensor): The real part of the complex tensor wit

Apply normalization to the input tensor. Args: x (Tensor): Input tensor with shape (batch, channels, ...). d

Modify the file path by replacing the input directory with the output directory, and optionally changing the file extension. Args:

Reads an audio file, resamples it to the target sampling frequency if necessary, and returns the audio waveform as a numpy array. Args:

Initilize PQMF module. The cutoff_ratio and beta parameters are optimized for #num_split = 4. See dicussion in https://github.com/kan

Initialize the WavehaxGenerator module. Args: in_channels (int): Number of conditioning feature channels. ch

dx: (B, N-1) or (N-1,) positive step sizes delta: (B, N-1) or (N-1,) slopes between points returns endpoint derivatives (d0, dn) sh

Write training loss metrics to TensorBoard. Args: loss (Dict): Dictionary containing loss metrics.

Perform complex convolution by combining the real and imaginary components. The complex convolution is defined as: - real_output = real_

Initialize weights for a complex-valued neural network layer. This function creates a complex-valued weight matrix with orthogonal initializ

Emulate np.round(x.astype(np.float32)) on GPU for non-negative x. Operates in float32 domain and returns float32.

Save the current state of the model to a checkpoint. Args: ckpt_path (str): Path to save the checkpoint file.

Validates and adjusts the lengths of feature arrays and corresponding audio data for alignment during audio processing. If audio data is prov

Initialize the ConvNeXtBlock2d module. Args: channels (int): Number of input and output channels for the block.

Initialize the STFT module. Args: n_fft (int): Number of Fourier transform points (FFT size). hop_length (in

Initialize the AdaptiveConv1d module. Args: in_channels (int): Number of input channels. out_channels (int):

Initialize the AdversarialLoss module. Args: average_by_discriminators (bool, optional): If True, the loss is averaged o

Check if it's time to evaluate the model.

Assert the audio and feature lengths are correctly adjusted for upsampling.

Check if it's time to log training metrics.

Check if it's time to save a checkpoint.

Check if training should be finished.

Evaluate model one epoch.

Perform a single evaluation step on a given batch. Args: batch (Tuple): A tuple containing batched data.

Batched path for B > 1. Strategy: - pack voiced x,y per batch into padded tensors of shape (B, maxNv) - compute

Fast path for B==1. This mirrors the exact SciPy-matching single-sample version you already validated.

Generate and save intermediate results during training. This method performs intermediate result visualization and logging during the trainin

Train the model for one complete epoch.

Perform a single training step on a given batch. Args: batch (Tuple): A tuple containing batched data.

Decompose the input signal into num_split subscale signals. Args: x (Tensor): Input signal with shape (B, 1, T).

Check if the filenames in two lists (without extensions) are identical. Args: list1 (List[str]): First list of file paths or names.

Compute and save statistics (mean and variance) for normalization of acoustic features. This function processes each feature specified in th

Converts an F0 sequence with intermittent zero values into a continuous F0 array by linearly interpolating over non-zero values. Args:

Design a prototype filter for Pseudo Quadrature Mirror Filter (PQMF) banks. Reference: - "A Kaiser window approach for the design of

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). This is the same as the DropConnect impl I created

Create a list file containing paths to feature files. Args: audio_scp (str): Path to the SCP file that lists input audio files.

Calculate forward propagation. Args: real (Tensor): Real part of the input tensor with shape (batch, channels, length).

Single-level forward DWT using Haar wavelet. Args: x (Tensor): Input tensor of shape (B, C, T). Returns:

Apply layer normalization to the input tensor. Args: x (Tensor): Input tensor with shape (batch, channels, height, width

Applies the inverse DWT to reconstruct a signal from its components. Args: low (Tensor): Low-frequency component.

Load the model state from a checkpoint. Args: ckpt_path (str): Path to the checkpoint file to be loaded. load_only_pa

Run the training process based on the provided configuration. Args: cfg (DictConfig): Configuration parameters for training.

Run the decoding process to generate audio waveforms from acoustic features. This function: - Loads a pre-trained GAN-based vocoder mode

Profile model parameters and MACs.

Create directories and prepare paths for feature files. Args: audio_paths (List[str]): List of input audio file paths. in_di

Perform pitch-dependent indexing for temporal sequences. This function applies pitch-dependent dilation to the input tensor `x`, retrieving

Perform pitch-dependent indexing for time-frequency feature maps. This function retrieves elements from a time-frequency map using pitch-dep

Reads a YAML configuration file and returns its contents as a dictionary. Args: file_path (str): Path to the YAML file. Returns

Execute the training loop.

Reconstruct the full-scale signal from subscale signals. Args: xs (List[Tensor]): List of subscale signals, each with sh

Convert real and imaginary components of a complex signal to log-magnitude and phase. Args: real (Tensor): Real part of the complex

Convert log-magnitude and implicit phase wrapping back to real and imaginary components of a complex signal. Args: log_magnitude (Te

Writes a dataset to an HDF5 file, optionally overwriting existing datasets. Args: hdf5_name (str): HDF5 file path. hdf5_path

Convert a list of audio-feature pairs into batched tensors. Args: batch (List): List of tuples containing audio, features, and F0

Get the specified index data. Args: idx (int): Index of the item. Returns: list: [filename (optiona

Get the specified index data. Args: idx (int): Index of the item. Returns: list: [filename (optiona

Initialize the ComplexConv2d module. Args: in_channels (int): Number of input channels. out_channels (int):

Initialize the ComplexActivation module. Args: activation (ModuleType): The activation function to apply to the real and

Initialize the ComplexNormLayer module. Args: channels (int): Number of channels (features) in the input. ep

Initialize the ComplexBatchNorm1d module. Args: channels (int): Number of input channels. eps (float, option

Initialize the ComplexLayerNorm1d module. Args: channels (int): Number of input channels (features). framewi

Initialize the ComplexBatchNorm2d module. Args: channels (int): Number of input channels. eps (float, option

Initialize the ComplexLayerNorm2d module. Args: channels (int): Number of input channels. framewise (bool, o

Initialize the ComplexConvNeXtBlock2d module. Args: channels (int): Number of input and output channels for the block.

Initialize the MelSpectrogram module. Args: sample_rate (int): Sampling frequency of input waveforms. hop_le

Initilize MultiStream1d module. Args: num_split (int): Number of num_split to decompose the input signal into. taps (

Initialize the DWT1d module. Args: num_split (int): Number of subscales (must be a power of two).

Initialize the LayerNorm1d module. Args: channels (int): Number of input features. eps (float, optional): A

Initialize the BatchNorm1d module. Args: channels (int): Number of input features. eps (float, optional): A

Initialize the LayerNorm2d module. Args: channels (int): Number of input features. framewise (bool, optional

Initialize the BatchNorm2d module. Args: channels (int): Number of input features. eps (float, optional): A

Initialize AdaptiveConv2d module. Args: in_channels (int): Number of input channels. out_channels (int): Num

Initialize the HiFiGANMultiPeriodDiscriminator module. Args: periods (List[int]): List of periods for each period discri

Initilize the SpectralDiscriminator module. Args: fft_size (int): Number of Fourier transform points for STFT.

Initilize the UnivNetMultiResolutionDiscriminator module. Args: fft_sizes (List[int]): List of FFT sizes for each spectr

Initilize the MultiResolutionMultiPeriodDiscriminator module. Args: periods (List[int]): List of periods for the HiFi-GA

Initialize the ComplexWavehaxGenerator module. Args: in_channels (int): Number of conditioning feature channels.

Initialize the MultiScaleWavehaxGenerator module. Args: in_channels (int): Number of conditioning feature channels.

Initialize the AudioFeatDataset. Args: sample_rate (int): Sampling frequency of the audio. hop_length (int):

Initialize the FeatDataset. Args: sample_rate (int): Sampling frequency of the audio. hop_length (int): Hop

Initialize the MelSpectralLoss module. Args: sample_rate (int): Sampling frequency of input waveforms. hop_l

Initialize the FeatureMatchingLoss module. Args: average_by_layers (bool, optional): If True, the loss is averaged over

Initialize Trainer. Args: cfg (dict): Configuration dictionary loaded from a YAML file. steps (int): Initial count of

Initialize the customized collater. Args: sample_rate (int): Sampling frequency. hop_length (int): Hop size for auxil

Return the number of items in the dataset. Returns: int: Dataset size.

Return the number of items in the dataset. Returns: int: Dataset size.

Calculate hinge loss for generator.