E3TTSOptions
Options & Configuration · AiDotNet.TextToSpeech.FlowDiffusion
Options for Gao et al.’s raw-waveform E3 TTS diffusion model.
Inheritance: ModelOptions → TtsModelOptions → EndToEndTtsOptions → E3TTSOptions
How It Works
The block arrays are Table 1 of the paper, flattened from block groups 0 through 3. E3 operates at 24 kHz and reduces a 10.92-second waveform by 4096x before the lowest-resolution attention blocks.
Constructors
| Constructor | Summary |
|---|---|
E3TTSOptions() |
Creates an instance with the default settings. |
E3TTSOptions(E3TTSOptions other) |
Initializes an independent copy. |
Properties
| Property | Summary |
|---|---|
bool FreezeBert { get; set; } |
The paper does not say that the English BERT checkpoint is fine-tuned; by default it is a frozen feature extractor. |
double BertDropout { get; set; } |
Gets or sets the dropout probability for the BERT. |
double DropoutRate { get; set; } |
Dropout probability in the BERT text encoder. |
double GroupNormEpsilon { get; set; } |
Gets or sets the numerical stability constant for group norm. |
double MinimumAlphaBar { get; set; } |
Terminal cumulative alpha in equation 7. |
double PaddingLossWeight { get; set; } |
Paper section 3.3 weights padded samples one tenth as much as real samples. |
double SilenceMagnitudeThreshold { get; set; } |
Gets or sets the silence magnitude threshold. |
double TrainingBetaEnd { get; set; } |
Last beta of the linear training schedule, separate from the Eq. |
double TrainingBetaStart { get; set; } |
First beta of the linear training schedule (Eq. |
int BertHeads { get; set; } |
Gets or sets the number of attention heads in the BERT. |
int BertHiddenDim { get; set; } |
Gets or sets the feature width of the BERT hidden. |
int BertIntermediateDim { get; set; } |
Gets or sets the feature width of the BERT intermediate. |
int BertLayers { get; set; } |
Gets or sets the number of layers in the BERT. |
int BertMaxPositions { get; set; } |
Gets or sets the BERT max positions. |
int BertVocabularySize { get; set; } |
Gets or sets the BERT vocabulary size. |
int ConditionDim { get; set; } |
Gets or sets the feature width of the condition. |
int DefaultSpeakerId { get; set; } |
Gets or sets the token ID for default speaker. |
int GroupNormGroups { get; set; } |
Gets or sets the group norm groups. |
int HiddenDim { get; set; } |
Gets the peak U-Net channel width. |
int InferenceWaveformSamples { get; set; } |
Fixed inference canvas; trailing low-magnitude blocks are removed after sampling. |
int NumEncoderLayers { get; set; } |
Number of layers in the BERT text encoder. |
int NumHeads { get; set; } |
Attention heads in the BERT text encoder. |
int SamplingSeed { get; set; } |
Gets or sets the random seed used during sampling. |
int SilenceBlockSize { get; set; } |
Gets or sets the silence block size. |
int SpeakerCount { get; set; } |
Gets or sets the speaker count. |
int TimeEmbeddingDim { get; set; } |
Gets or sets the feature width of the time embedding. |
int TrainingWaveformSamples { get; set; } |
10.92 seconds at 24 kHz, as used during training. |
int[] AdaptiveKernelCounts { get; set; } |
Number of softmax kernel experts; zero means the block uses FiLM instead. |
int[] AttentionHeads { get; set; } |
Self/cross-attention heads per block; zero means no attention. |
int[] BlockChannels { get; set; } |
Output width of each DBlock in Table 1. |
int[] BlockKernelSizes { get; set; } |
Convolution kernel of each DBlock/UBlock pair in Table 1. |
int[] BlockRepetitions { get; set; } |
Residual blocks at each resolution, from Table 1’s Blocks row. |
int[] BlockStrides { get; set; } |
Temporal stride of each DBlock/UBlock pair in Table 1; the product is 4096. |