WaveRNNOptions

Options & Configuration · AiDotNet.TextToSpeech.Vocoders

Options for WaveRNN (Kalchbrenner et al. 2018): a single-layer recurrent vocoder with a dual softmax over the coarse and fine bytes of 16-bit samples, optionally sparsified by magnitude pruning.

Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → WaveRNNOptions

For Beginners

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

How It Works

Defaults are the paper’s WaveRNN-896 (§2, §5): a state of 896 units split into coarse and fine halves; 24 kHz 16-bit audio; training on 960-sample sequences with full back-propagation through time. Weight pruning (§3) is off (SparsityTarget 0); when on, the recurrent matrices are pruned by magnitude every 500 steps from step 1,000 over 200,000 steps along z = Z(1 − (1 − (t − t₀)/S)³), optionally in 16×1 or 4×4 blocks.

What the paper leaves open: it conditions on linguistic features and pitch without saying how, so the mel spectrogram (Tacotron 2’s 80 bands, 50 ms window, 12.5 ms hop, 125 Hz–7.6 kHz) is upsampled by a transposed convolutional network as in WaveNet and projected into the gates; the optimizer follows fatchord/WaveRNN (Adam at 1e-4, batch 32, gradient norm clipped to 4).

Constructors

Constructor Summary
WaveRNNOptions() Creates the paper’s WaveRNN-896 configuration.
WaveRNNOptions(WaveRNNOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
double MaxGradientNorm { get; set; } Gets or sets the global gradient-norm clip (4).
double MelMaxFrequency { get; set; } Gets or sets the highest mel frequency (7.6 kHz).
double MelMinFrequency { get; set; } Gets or sets the lowest mel frequency (125 Hz).
double SparsityTarget { get; set; } Gets or sets the final fraction Z of pruned recurrent weights (0: the dense WaveRNN).
int HiddenDim { get; set; } The actual recurrent state width, split into equal coarse and fine halves.
int HopSize { get; set; } The product of the actual conditioning upsampling factors.
int PruneBlockColumns { get; set; } Gets or sets the columns of a pruning block (1; 4 for the 4×4 structure).
int PruneBlockRows { get; set; } Gets or sets the rows of a pruning block (1; 16 for the 16×1 structure, 4 for 4×4).
int PruningInterval { get; set; } Gets or sets the steps between mask updates (500).
int PruningStartStep { get; set; } Gets or sets the step t₀ at which pruning begins (1,000).
int PruningSteps { get; set; } Gets or sets the pruning steps S over which sparsity reaches the target (200,000).
int RnnDim { get; set; } Gets or sets the state size (896), split into coarse and fine halves.
int SamplingSeed { get; set; } Gets or sets the seed of the synthesis draws.
int SequenceSamples { get; set; } Gets or sets the exact samples of a training sequence (960; full back-propagation through time).
int WindowSize { get; set; } Gets or sets the analysis window of the input features (1200: 50 ms).
int[] UpsampleScales { get; set; } Gets or sets the conditioning upsampler’s factors (5, 5, 4, 3; their product is the hop).