WaveGlowOptions
Options & Configuration · AiDotNet.TextToSpeech.Vocoders
Options for WaveGlow (Prenger et al. 2019): a flow-based vocoder trained only by maximum likelihood.
Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → WaveGlowOptions
For Beginners
These options configure the WaveGlow model. Default values follow the original paper settings.
How It Works
Defaults are the paper’s (§2–3): groups of 8 samples; 12 steps of an invertible 1×1 convolution and an affine coupling; 2 channels output early after every 4 couplings; WN networks of 8 weight-normalized dilated convolutions with 3 taps, 512 residual and 256 skip channels; σ = √0.5 in training and 0.6 at inference; 80-band mel spectrograms (librosa’s Slaney filters) from a 1024-point FFT with a 1024-sample window and hop 256 at 22.05 kHz; Adam at 1e-4, batch 24, 16,000-sample clips (the paper lowers the rate to 5e-5 by hand when training plateaus).
What the paper leaves open follows NVIDIA/waveglow: the mel spectrogram upsampled by a transposed convolution
with a 1024-tap kernel and stride 256; Tacotron 2’s features (centred STFT, ln(max(x, 1e-5)), 0–8 kHz); the
gated units as wide as the residual channels (the paper names only residual and skip widths); no gradient clipping;
the loss (Σz² / 2σ² − Σ log s − Σ log|det W|) / samples.
Constructors
| Constructor | Summary |
|---|---|
WaveGlowOptions() |
Creates the paper’s WaveGlow configuration. |
WaveGlowOptions(WaveGlowOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double InferenceSigma { get; set; } |
Gets or sets σ of the latent Gaussian at inference (0.6). |
double MelMaxFrequency { get; set; } |
Gets or sets the highest mel frequency (8 kHz). |
double MelMinFrequency { get; set; } |
Gets or sets the lowest mel frequency (0 Hz). |
double TrainingSigma { get; set; } |
Gets or sets σ of the latent Gaussian in training (√0.5). |
int EarlyOutputChannels { get; set; } |
Gets or sets the channels output early each time (2). |
int EarlyOutputEvery { get; set; } |
Gets or sets the couplings between early outputs (4). |
int GateChannels { get; set; } |
Gets or sets the WN gated-unit channels (512). |
int GroupSize { get; set; } |
Gets or sets the samples per squeezed vector (8). |
int KernelSize { get; set; } |
Gets or sets the WN convolution taps (3). |
int NumFlows { get; set; } |
Gets or sets the steps of flow (12). |
int NumWaveNetLayers { get; set; } |
Gets or sets the WN layers per coupling (8). |
int ResidualChannels { get; set; } |
Gets or sets the WN residual channels (512). |
int SamplingSeed { get; set; } |
Gets or sets the seed of the training draws and of the synthesis latents. |
int SegmentSamples { get; set; } |
Gets or sets the samples of a training clip (16,000; rounded down to whole frames). |
int SkipChannels { get; set; } |
Gets or sets the WN skip channels (256). |
int UpsampleKernel { get; set; } |
Gets or sets the mel upsampler’s kernel (1024). |
int WindowSize { get; set; } |
Gets or sets the analysis window of the input features (1024). |