StyleTTSZSOptions
Options & Configuration · AiDotNet.TextToSpeech.StyleEmotion
Options for StyleTTSZS.
Inheritance: ModelOptions → StyleTtsBaseOptions → StyleTtsWaveformOptions → StyleTTSZSOptions
For Beginners
These settings size StyleTTS-ZS’s networks and control how it samples the speaking style of a short prompt recording.
How It Works
The defaults are the paper’s (Li et al. 2024, §III–IV and Appendix C–D): 512-wide conformer modules (8 heads of 64, feed-forward 1024; Tables VII–XI), a 50-row time-varying style quantized by 9 codebooks of 1024 codes in 8 dimensions, the style denoiser’s 100-step DDIM sampling with guidance 5, the distilled student, AdamW (β = 0, 0.99, decay 1e-4, rate 1e-4). StyleTTS-ZS released no code: where the paper is silent the values follow StyleTTS 2’s code (the decoder, aligner, pitch extractor, PL-BERT and wave discriminators it says it reuses) or the cited components (wenet’s conformer, DAC’s quantizer, WeSpeaker’s ResNet34), each noted below.
Constructors
| Constructor | Summary |
|---|---|
StyleTTSZSOptions() |
Creates the paper’s options. |
StyleTTSZSOptions(StyleTTSZSOptions other) |
Copies another instance. |
Properties
| Property | Summary |
|---|---|
bool Separators { get; set; } |
Whether a learned separator row sits between joined sequences (the |
bool UseDistilledSampler { get; set; } |
Whether inference samples with the distilled one-step student (true) or the teacher’s DDIM. |
double BertBeta1 { get; set; } |
PL-BERT’s AdamW β₁ (0.9, StyleTTS 2’s). |
double BertLearningRate { get; set; } |
PL-BERT’s learning rate (1e-5, StyleTTS 2’s; the paper is silent). |
double BertWeightDecay { get; set; } |
PL-BERT’s AdamW weight decay (0.01, StyleTTS 2’s). |
double CodebookWeight { get; set; } |
Weight of the quantizer’s codebook loss (1, DAC’s). |
double CommitmentWeight { get; set; } |
Weight of the quantizer’s commitment loss (0.25, DAC’s; the paper lists no quantizer loss). |
double ConditionDropout { get; set; } |
Rate at which training drops the denoiser’s conditions (0.1). |
double ConformerDropout { get; set; } |
Dropout inside the conformer blocks (0.1, wenet’s default; the paper is silent). |
double DistillationGuidanceMax { get; set; } |
Largest guidance scale distillation draws (15, Eq. |
double DistillationGuidanceMin { get; set; } |
Smallest guidance scale distillation draws (1, Eq. |
double F0LossWeight { get; set; } |
λ_f0 (0.1, App. |
double Guidance { get; set; } |
Classifier-free guidance scale ω at inference (5). |
double MonotonicLossWeight { get; set; } |
λ_mono (5, App. |
double NormLossWeight { get; set; } |
λ_n (1, App. |
double ProsodyMaskProbability { get; set; } |
Probability that a training step masks a random span of the prosody encoder’s input (0.5; the paper states the masking without its rate). |
double S2sLossWeight { get; set; } |
λ_s2s (0.2, App. |
double StyleDropout { get; set; } |
Dropout on the global style in training (0.2, §III-A). |
int CodeDim { get; set; } |
Dimension of the factorized codes (8, following DAC). |
int CodebookSize { get; set; } |
Codes per codebook (1024). |
int FeedForwardDim { get; set; } |
Feed-forward width of the conformer blocks (1024). |
int Heads { get; set; } |
Attention heads of the conformer blocks (8, of 64 features each). |
int HopSize { get; set; } |
Mel hop paired with the decoder rates and, for iSTFTNet, the inverse-STFT hop. |
int MaxPhonemes { get; set; } |
Phonemes the prosody encoder’s position embedding covers (512, PL-BERT’s positions). |
int ModelDim { get; set; } |
Width of every new module (512, Tables VII–XI). |
int PromptFrames { get; set; } |
Mel frames of the longest training prompt, a random segment of the utterance (240: 3 s, as the evaluation prompts; the paper truncates prompts to the batch’s shortest utterance). |
int Quantizers { get; set; } |
Codebooks of the residual quantizer (9). |
int SamplingSteps { get; set; } |
DDIM steps L of the teacher’s sampling (100). |
int SpeakerChannels { get; set; } |
First width of the WeSpeaker ResNet of the speaker-embedding feature matching (32: ResNet34). |
int SpeakerEmbedding { get; set; } |
Its embedding width (256). |
int StyleDim { get; set; } |
Global style width, equal to the native prompt-text encoder’s ModelDim. |
int StyleLength { get; set; } |
Rows K of the time-varying style (50). |
int[] DenoiserKernels { get; set; } |
Kernels of the style denoiser’s and the student’s conformer blocks (two of 31, ten of 15; Table X). |
int[] DiscriminatorKernels { get; set; } |
Kernels of each multimodal discriminator’s conformer blocks (six of 15; Table XI). |
int[] PromptTextKernels { get; set; } |
Kernels of the prompt-text encoder’s conformer blocks (one of 31, five of 15; Table VII). |
int[] ProsodyDecoderKernels { get; set; } |
Kernels of each prosody decoder’s conformer blocks (six of 31; Table IX). |
int[] ProsodyEncoderKernels { get; set; } |
Kernels of the prosody encoder’s conformer blocks (one of 31, five of 15; Table VIII). |
int[] SpeakerBlocks { get; set; } |
Residual blocks per stage of that ResNet ({3, 4, 6, 3}: ResNet34). |
int[] UpsampleKernels { get; set; } |
One kernel per upsampling stage, twice the corresponding positive rate. |
int[] UpsampleRates { get; set; } |
Decoder upsampling factors paired with doubled kernels and the mel hop. |