DiffWaveOptions
Options & Configuration · AiDotNet.TextToSpeech.Vocoders
Options for DiffWave (Kong et al. 2021): a diffusion vocoder whose noise predictor is a non-autoregressive, bidirectional dilated-convolution network conditioned on the mel spectrogram.
Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → DiffWaveOptions
Known derived types: PriorGradOptions
For Beginners
These options configure the DiffWave model. Default values follow the original paper settings.
How It Works
Defaults are the paper’s DiffWave BASE with T = 50 (§5.1, Table 1): 30 residual layers of C = 64 channels, kernel 3,
dilation cycle 1, 2, …, 512; a linear β schedule from 1e-4 to 0.05 over 50 steps; 22.05 kHz audio, 80 mel bands
from a 1024-point FFT with hop 256; the mel upsampled ×256 by two transposed 2-D convolutions of stride 16 and
(3, 32) kernels with leaky ReLU 0.4; Adam at 2e-4, batch 16, about 16,000-sample clips; the fast-sampling schedule
1e-4, 1e-3, 0.01, 0.05, 0.2, 0.5 (App. B; set UseFastSampling).
What the paper leaves open follows lmnt-com/diffwave (params.py, preprocess.py): 62-frame crops;
mel features from torchaudio (centred, window-normalized magnitude, HTK bands from 20 Hz to Nyquist, a 1024-sample
window) as clamp((20 log10(max(x, 1e-5)) − 20 + 100) / 100, 0, 1); each sampling step clamped to [−1, 1]; no
gradient clipping. The paper’s L2 noise loss is used (the reference’s code uses L1).
The paper’s other two tasks have their own configurations: Unconditional (§5.2) and
ClassConditional (§5.3). Both generate whole utterances (the paper trains on full one-second SC09
clips), so a model input is the class label (or, unconditionally, ignored) rather than a spectrogram.
Constructors
| Constructor | Summary |
|---|---|
DiffWaveOptions() |
Creates the paper’s DiffWave BASE (T = 50) configuration. |
DiffWaveOptions(DiffWaveOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
DiffWaveConditioner Conditioner { get; set; } |
Gets or sets what the noise predictor is conditioned on (the mel spectrogram). |
bool UseFastSampling { get; set; } |
Gets or sets whether synthesis uses the fast-sampling schedule (false: the training schedule). |
double MelMinFrequency { get; set; } |
Gets or sets the lowest mel frequency (20 Hz). |
double[] InferenceNoiseSchedule { get; set; } |
Gets or sets the fast-sampling β schedule (1e-4, 1e-3, 0.01, 0.05, 0.2, 0.5). |
double[] NoiseSchedule { get; set; } |
Gets or sets the training β schedule (50 values linearly spaced from 1e-4 to 0.05). |
int CropFrames { get; set; } |
Gets or sets the mel frames of a training crop (62). |
int DilationCycle { get; set; } |
Gets or sets the dilation cycle length (10: dilations 1 … 512). |
int HiddenDim { get; set; } |
The actual residual denoiser width (64 for vocoding, 256 for the paper’s utterance models). |
int HopSize { get; set; } |
The product of the mel conditioner’s even transposed-convolution strides. |
int NumClasses { get; set; } |
Gets or sets the number of labels of the class conditioner (no default: it is the dataset’s; SC09 has 10 digits). |
int NumResLayers { get; set; } |
Gets or sets the residual layers N (30). |
int ResChannels { get; set; } |
Gets or sets the residual channels C (64 for BASE; 128 for LARGE). |
int SamplingSeed { get; set; } |
Gets or sets the seed of the training draws and of the synthesis noise. |
int UtteranceSamples { get; set; } |
Gets or sets the length L of a generated utterance without a spectrogram (16,000 in §5.2). |
int WindowSize { get; set; } |
Gets or sets the analysis window of the input features (1024). |
int[] UpsampleStrides { get; set; } |
Gets or sets the spectrogram upsampler’s strides (16, 16: their product is the hop). |
Methods
| Method | Summary |
|---|---|
DiffWaveOptions ClassConditional(int numClasses) |
The paper’s class-conditional configuration (§5.3): the unconditional model’s hyperparameters with a shared 128-wide embedding of one of numClasses labels (10 digits for SC09) added in every residual layer. |
DiffWaveOptions Unconditional() |
The paper’s unconditional generation configuration (§5.2): 16 kHz one-second utterances (L = 16,000), 36 residual layers of C = 256 channels, dilation cycle 1, 2, …, 2048, T = 200 steps with β linear from 1e-4 to 0.02, sampled over the full… |
double[] Linear(double from, double to, int count) |
Values linearly spaced from from to to (numpy linspace). |