TortoiseTTSOptions
Options & Configuration · AiDotNet.TextToSpeech.CodecBased
Options for TorToiSe: Better Speech Synthesis Through Scaling (Betker, 2023).
Inheritance: ModelOptions → TtsModelOptions → CodecTtsOptions → TortoiseTTSOptions
How It Works
Defaults reproduce the released standard preset. Smaller values may be supplied for experiments and tests, but the topology is unchanged: VQ-VAE codes, a GPT-2 prior, CLVP candidate ranking, a latent-conditioned diffusion decoder, and the 24 kHz UnivNet-c32 vocoder.
Constructors
| Constructor | Summary |
|---|---|
TortoiseTTSOptions() |
Creates the paper/release configuration. |
Properties
| Property | Summary |
|---|---|
TortoiseTrainingComponent TrainingComponent { get; set; } |
The component optimized by the TTS training entry points. |
UnivNetOptions VocoderTraining { get; set; } |
UnivNet’s original auxiliary STFT, discriminator and pretraining settings, used only by the vocoder stage. |
double ClassifierFreeGuidanceScale { get; set; } |
Released classifier-free guidance k (2), ramped over the selected DDPM schedule. |
double ClvpTokenDropout { get; set; } |
Token dropout used while training CLVP (15%). |
double DiffusionConditioningDropout { get; set; } |
Probability of dropping all diffusion conditioning during training (15%). |
double RepetitionPenalty { get; set; } |
Penalty applied to a code already present in the generated sequence (2). |
double Temperature { get; set; } |
Autoregressive sampling temperature (0.8). |
double TextLossWeight { get; set; } |
Weight of the text cross-entropy beside the speech-code loss (0.01). |
double TokenizerCommitmentWeight { get; set; } |
VQ-VAE commitment-loss multiplier. |
double TopP { get; set; } |
Nucleus probability used by the paper’s autoregressive sampler (0.8). |
double[] AutoregressiveMelNormalization { get; set; } |
Optional checkpoint mel_norms values; null uses the released 80-band values. |
int AutoregressiveFeedForwardDim { get; set; } |
GPT-2 feed-forward width (4,096). |
int AutoregressiveFftSize { get; set; } |
FFT size used for 22.05 kHz conditioning features (1024). |
int AutoregressiveHopSize { get; set; } |
Hop used for 22.05 kHz conditioning features (256). |
int AutoregressiveMelChannels { get; set; } |
Mel bands read by the speech tokenizer and autoregressive conditioning encoder (80). |
int ClvpDim { get; set; } |
CLVP transformer/latent width (768). |
int ClvpHeads { get; set; } |
CLVP attention heads (12). |
int ClvpLayers { get; set; } |
CLVP transformer depth for each of its two towers (20). |
int ClvpSpeechLength { get; set; } |
CLVP maximum speech positions in the released checkpoint (430). |
int ClvpTextLength { get; set; } |
CLVP maximum text positions (350). |
int ConditioningSamples { get; set; } |
Maximum samples in one autoregressive conditioning clip (132,300, or six seconds). |
int DiffusionConditioningSamples { get; set; } |
Maximum 24 kHz reference samples read by the diffusion speaker encoder (102,400). |
int DiffusionDim { get; set; } |
Diffusion model width (1,024). |
int DiffusionHeads { get; set; } |
Diffusion attention heads (16). |
int DiffusionInferenceSteps { get; set; } |
Spaced DDPM steps in the released standard preset (200). |
int DiffusionLayers { get; set; } |
Alternating attention/convolution diffusion layers (10). |
int DiffusionTrainingSteps { get; set; } |
Training diffusion steps used by the release (4,000). |
int MaxConditioningInputs { get; set; } |
Maximum conditioning clips averaged by the released model (2). |
int MaximumAutoregressiveTokens { get; set; } |
Generation/padding limit in the released synthesis API (500), separate from the checkpoint’s 604-position capacity. |
int MinimumAutoregressiveTokens { get; set; } |
Minimum generated speech codes before a stop token may terminate a candidate. |
int NumAutoregressiveSamples { get; set; } |
Number of autoregressive candidates passed to CLVP (256 in the standard preset). |
int SamplingSeed { get; set; } |
Seed for candidate sampling, diffusion noise, and UnivNet noise. |
int SpeechStartToken { get; set; } |
Speech start token (8,192). |
int SpeechStopToken { get; set; } |
Speech stop token (8,193). |
int TextStartToken { get; set; } |
Text start token (255). |
int TextStopToken { get; set; } |
Text stop/padding token (0). |
int TokenizerCodeDim { get; set; } |
Released dVAE code-vector width (512). |
int TokenizerCompression { get; set; } |
Temporal compression of the tokenizer (4). |
int TokenizerLatentDim { get; set; } |
VQ-VAE latent channels (512). |
int TokenizerResidualBlocks { get; set; } |
Residual blocks in each dVAE encoder/decoder (3). |
int TokenizerSampleRate { get; set; } |
The VQ-VAE and autoregressive conditioner’s rate (22.05 kHz). |
int UnivNetChannels { get; set; } |
UnivNet-c32 generator channels (32). |
int UnivNetKernelPredictorConvSize { get; set; } |
UnivNet kernel-predictor convolution width (3). |
int UnivNetKernelPredictorHidden { get; set; } |
UnivNet kernel-predictor hidden channels (64). |
int UnivNetNoiseDim { get; set; } |
UnivNet input noise channels (64). |
int UnivNetPaddingFrames { get; set; } |
Log-mel frames appended by the released vocoder and cropped from its output (10). |
int[] UnivNetDilations { get; set; } |
UnivNet LVC dilations (1, 3, 9, 27). |
int[] UnivNetUpsampleRates { get; set; } |
UnivNet LVC upsampling rates (8, 8, 4; product 256). |
string TextTokenizerPath { get; set; } |
Optional original tokenizer.json path; null uses the released 255-token BPE resource. |