ParallelWaveGANOptions

Options & Configuration · AiDotNet.TextToSpeech.Vocoders

Options for Parallel WaveGAN (Yamamoto et al. 2020): a non-autoregressive WaveNet generator over Gaussian noise, trained with the multi-resolution STFT loss and an adversarial loss.

Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → ParallelWaveGANOptions

For Beginners

These options configure the Parallel WaveGAN model. Default values follow the original paper settings.

How It Works

Defaults are the paper’s (§4.1): 24 kHz audio; 80 log-mel bands from 70 to 8000 Hz with 50 ms frames and a 12.5 ms shift (1200 / 300 samples), standardized; a generator of 30 dilated residual blocks in three cycles with 64 residual and skip channels and kernel 3; a discriminator of ten 64-channel convolutions; λ_adv = 4; STFT resolutions (1024, 600, 120), (2048, 1200, 240), (512, 240, 50); 400k steps with the discriminator fixed for the first 100k; RAdam (ε = 1e-6) at 1e-4 (generator) and 5e-5 (discriminator), halved every 200k steps; batch 8 of 1-second clips.

What the paper leaves open follows kan-bayashi/ParallelWaveGAN (parallel_wavegan.v1.yaml): a 2048-point FFT for the features; a 128-channel gate; upsampling scales 4, 5, 3, 5 with an auxiliary context window of 2; gradient-norm clipping at 10 (generator) and 1 (discriminator).

Constructors

Constructor Summary
ParallelWaveGANOptions() Creates the paper’s configuration.
ParallelWaveGANOptions(ParallelWaveGANOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
double AdversarialWeight { get; set; } Gets or sets λ_adv (4).
double DiscriminatorGradientClipNorm { get; set; } Gets or sets the discriminator’s gradient-norm clip (1).
double DiscriminatorLearningRate { get; set; } Gets or sets the discriminator’s learning rate (5e-5).
double Epsilon { get; set; } Gets or sets RAdam’s ε (1e-6).
double GeneratorGradientClipNorm { get; set; } Gets or sets the generator’s gradient-norm clip (10).
double MelMaxFrequency { get; set; } Gets or sets the highest mel frequency (8000 Hz).
double MelMinFrequency { get; set; } Gets or sets the lowest mel frequency (70 Hz).
double[] MelMean { get; set; } Gets or sets the per-band mean of the log-mel features over the training set, or null for none.
double[] MelScale { get; set; } Gets or sets the per-band standard deviation of the log-mel features over the training set, or null for none.
int AuxContextWindow { get; set; } Gets or sets the mel frames of context on each side of the input convolution (2).
int DiscriminatorChannels { get; set; } Gets or sets the discriminator’s channels (64).
int DiscriminatorLayers { get; set; } Gets or sets the discriminator’s convolutions (10).
int GateChannels { get; set; } Gets or sets the gate channels (128).
int HopSize { get; set; } The product of the actual learned conditioning upsampler’s scales (300 in the paper).
int KernelSize { get; set; } Gets or sets the convolution kernel (3).
int LearningRateHalvingSteps { get; set; } Gets or sets the steps between learning-rate halvings (200,000).
int NumLayers { get; set; } Gets or sets the generator’s residual blocks (30).
int NumStacks { get; set; } Gets or sets the dilation cycles (3).
int ResidualChannels { get; set; } Gets or sets the residual channels (64).
int SamplingSeed { get; set; } Gets or sets the seed of the training draws and of the synthesis noise.
int SegmentSize { get; set; } Gets or sets the training segment in samples (24,000: one second).
int SkipChannels { get; set; } Gets or sets the skip channels (64).
int WindowSize { get; set; } Gets or sets the analysis window of the input features (1200: 50 ms).
int[] StftFftSizes { get; set; } Gets or sets the STFT loss FFT sizes (1024, 2048, 512).
int[] StftHopSizes { get; set; } Gets or sets the STFT loss hops (120, 240, 50).
int[] StftWindowSizes { get; set; } Gets or sets the STFT loss windows (600, 1200, 240).
long DiscriminatorStartStep { get; set; } Gets or sets the steps the discriminator stays fixed (100,000).