NonAttentiveTacotronOptions
Options & Configuration · AiDotNet.TextToSpeech.Classic
Options for Non-Attentive Tacotron (supervised-duration variant).
Inheritance: ModelOptions → TtsModelOptions → AcousticModelOptions → NonAttentiveTacotronOptions
For Beginners
These options configure the NonAttentiveTacotron model. Default values follow the original paper settings.
How It Works
Defaults are Appendix A of Shen et al. 2020: 24 kHz audio, 50 ms Hann frames (1200 samples), 12.5 ms hop (300), 2048-point FFT, 128 mel channels; token embedding 512; encoder convolutions [512, 512, 512] of kernel 5 with batch normalization (decay 0.999) and no activation, a 512 × 2 bidirectional LSTM; duration and range predictors of two 512 × 2 bidirectional LSTMs; positional embedding 32; decoder pre-net [256, 256] with ReLU and dropout 0.5, two 1024-unit LSTMs with zoneout 0.1 and cell cap 10; post-net [512, 512, 512, 512, K] of kernel 5; λ_dur = 2.0; reduction factor 2 (§4.1).
Constructors
| Constructor | Summary |
|---|---|
NonAttentiveTacotronOptions() |
Creates the paper’s configuration. |
NonAttentiveTacotronOptions(NonAttentiveTacotronOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double BatchNormDecay { get; set; } |
Gets or sets the encoder batch normalization decay (0.999). |
double DurationLossWeight { get; set; } |
Gets or sets λ_dur (2.0, supervised). |
double DurationScale { get; set; } |
Gets or sets the pace control applied to predicted durations (1; §4.2 divides durations by a factor). |
double LstmCellClip { get; set; } |
Gets or sets the LSTM cell absolute value cap (10). |
double PostnetDropout { get; set; } |
Gets or sets the post-net dropout (not stated; 0). |
double PrenetDropout { get; set; } |
Gets or sets the pre-net dropout (0.5). |
double ZoneoutProbability { get; set; } |
Gets or sets the LSTM zoneout probability (0.1). |
int DecoderDim { get; set; } |
Gets or sets the autoregressive decoder LSTM width. |
int DecoderLstmDim { get; set; } |
Gets or sets the decoder LSTM width (1024). |
int DurationLstmDim { get; set; } |
Gets or sets the duration predictor’s bidirectional LSTM width per direction (512). |
int EmbeddingDim { get; set; } |
Gets or sets the token embedding width (512). |
int EncoderDim { get; set; } |
Gets or sets the encoder LSTM width per direction. |
int EncoderKernelSize { get; set; } |
Gets or sets the encoder convolutions’ kernel (5). |
int EncoderLstmDim { get; set; } |
Gets or sets the encoder bidirectional LSTM width per direction (512). |
int HiddenDim { get; set; } |
Gets or sets the native token embedding width. |
int NumEncoderLayers { get; set; } |
Gets or sets the encoder convolution count, extending the last configured width when needed. |
int PositionalEmbeddingDim { get; set; } |
Gets or sets the positional embedding width (32). |
int PostnetKernelSize { get; set; } |
Gets or sets the post-net kernel (5). |
int RangeLstmDim { get; set; } |
Gets or sets the range predictor’s bidirectional LSTM width per direction (512). |
int WindowSize { get; set; } |
Gets or sets the analysis window in samples (50 ms at 24 kHz: 1200). |
int[] EncoderConvChannels { get; set; } |
Gets or sets the encoder convolutions’ channels ([512, 512, 512]). |
int[] PrenetSizes { get; set; } |
Gets or sets the decoder pre-net widths ([256, 256], supervised). |