TacotronOptions

Options & Configuration · AiDotNet.TextToSpeech.Classic

Options for Tacotron (attention-based sequence-to-sequence TTS with CBHG modules).

Inheritance: ModelOptions → TtsModelOptions → AcousticModelOptions → TacotronOptions

For Beginners

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

How It Works

Defaults are the paper’s (Wang et al. 2017, §4, Table 1): 24 kHz audio, 50 ms Hann frames (1200 samples) with a 12.5 ms shift (300), 2048-point FFT and 0.97 pre-emphasis; 256-wide character embedding; pre-nets FC-256 and FC-128 with ReLU and dropout 0.5; an encoder CBHG with K = 16 and 128-wide convolutions, highways and GRUs; a 256-cell attention GRU, 2 residual 256-cell decoder GRUs and r = 2 frames per step; a post-processing CBHG with K = 8 and projections of 256 and 80; Griffin–Lim with 50 iterations on magnitudes raised to 1.2; Adam from 0.001.

Constructors

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

Properties

Property Summary
double GriffinLimPower { get; set; } Gets or sets the power the predicted magnitudes are raised to before Griffin–Lim (1.2).
double PreEmphasis { get; set; } Gets or sets the pre-emphasis coefficient (0.97).
double PrenetDropout { get; set; } Gets or sets the pre-nets’ dropout (0.5).
int AttentionDim { get; set; } Gets or sets the attention GRU and attention width (256).
int CbhgChannels { get; set; } Gets or sets the width of the CBHG bank convolutions, highways and each GRU direction (128).
int DecoderRnnDim { get; set; } Gets or sets the decoder GRUs’ width (256).
int EmbeddingDim { get; set; } Gets or sets the character embedding width (256).
int EncoderBankSize { get; set; } Gets or sets the encoder CBHG’s convolution bank size K (16).
int GriffinLimIterations { get; set; } Gets or sets the Griffin–Lim iterations (50).
int MaxDecoderSteps { get; set; } Gets or sets the decoder steps synthesis runs (the paper states no stopping rule; the reference implementation decodes a fixed 1000 frames, which is 500 steps at r = 2).
int PostBankSize { get; set; } Gets or sets the post-processing CBHG’s convolution bank size K (8).
int PostProjectionChannels { get; set; } Gets or sets the post-processing CBHG’s first projection width (256).
int WindowSize { get; set; } Gets or sets the analysis window in samples (50 ms at 24 kHz: 1200).
int[] PrenetSizes { get; set; } Gets or sets the widths of the encoder and decoder pre-nets’ two layers (256, 128).