MegaTTS3Options

Options & Configuration · AiDotNet.TextToSpeech.Latest

Options for MegaTTS3: the release’s five networks (the diffusion transformer, the duration language model, the Whisper aligner, the Qwen2 grapheme-to-phoneme model and the WaveVAE), its inference settings and the paper’s training recipe.

Inheritance: ModelOptions → TtsModelOptions → MegaTTS3Options

For Beginners

MegaTTS 3 copies a voice from a few seconds of it. These options size its parts and set how it samples; the defaults load the released checkpoints.

How It Works

The defaults are the release’s (bytedance/MegaTTS3: dit.py, duration_lm/config.yaml, whisper_small.py, g2p/config.json, wavvae_v3.py, infer_cli.py) and, for training, the paper’s (Jiang et al. 2025, §3 and Appendix A.1).

Constructors

Constructor Summary
MegaTTS3Options() Initializes a new instance with default values.

Properties

Property Summary
bool SparseAlignment { get; set; } Whether the transformer reads one random anchor frame per phone in training (the paper’s sparse alignment); false reads every frame, as inference does.
double Beta1 { get; set; } Adam’s β1 (0.9).
double Beta2 { get; set; } Adam’s β2 (0.999).
double DropoutRate { get; set; } The released networks have no model dropout; only zero is supported.
double DurationDisturbance { get; set; } The random disturbance of predicted durations (dur_disturb, 0.1).
double DurationScale { get; set; } The scale of predicted durations (dur_alpha, 1.0).
double G2pRopeTheta { get; set; } The g2p model’s RoPE base (1,000,000).
double PromptFractionMax { get; set; } See PromptFractionMin.
double PromptFractionMin { get; set; } The prompt’s share of a training latent, drawn from U(min, max) (0.1, 0.9).
double SpeakerDropProbability { get; set; } The probability of dropping the speaker prompt (p_spk, 0.10).
double SpeakerGuidance { get; set; } The speaker guidance weight (--t_w, “similarity”, 2.5).
double TextDropProbability { get; set; } The probability of also dropping the text when the prompt is dropped (50 %).
double TextGuidance { get; set; } The text guidance weight (--p_w, “intelligibility”, 1.6).
int AlignerHeads { get; set; } The aligner’s heads (8).
int AlignerLayers { get; set; } The aligner’s encoder and decoder layers (6 each).
int AlignerMaxTokens { get; set; } The aligner’s most decoded tokens (768).
int AlignerVocabulary { get; set; } The aligner’s vocabulary (6800).
int AlignerWidth { get; set; } The aligner’s width (512).
int ChineseChunkLimit { get; set; } Chinese chunks of about this many ideographs (60).
int ContextMaskDim { get; set; } The context mask’s projection width (16).
int DitHeads { get; set; } Attention heads (16).
int DitLayers { get; set; } Transformer layers (24).
int DitPhoneVocabulary { get; set; } The transformer’s phone vocabulary (302; the last is the CFG mask token).
int DitToneVocabulary { get; set; } The transformer’s tone vocabulary (32; the last is the CFG mask token).
int DurationCodeSize { get; set; } Duration codes (128: a phone lasts at most 127 frames).
int DurationEncoderFilter { get; set; } The text encoder’s feed-forward width (1024).
int DurationEncoderHeads { get; set; } The text encoder’s heads (2).
int DurationEncoderLayers { get; set; } The text encoder’s layers (4).
int DurationEncoderWidth { get; set; } The duration model’s text encoder width (512).
int DurationHeads { get; set; } The duration decoder’s heads (8).
int DurationLayers { get; set; } The duration decoder’s layers (8).
int DurationPhoneVocabulary { get; set; } The duration model’s phone vocabulary (174: the release’s 171 phones and three reserved tokens).
int DurationToneVocabulary { get; set; } The duration model’s tone vocabulary (23: 20 and three reserved).
int DurationWidth { get; set; } The duration decoder’s width (512).
int EnglishChunkBytes { get; set; } English chunks of at most this many UTF-8 bytes (130).
int G2pHeads { get; set; } The g2p model’s attention heads (14).
int G2pHidden { get; set; } The g2p model’s width (896).
int G2pIntermediate { get; set; } The g2p model’s feed-forward width (4864).
int G2pKeyValueHeads { get; set; } The g2p model’s key-value heads (2).
int G2pLayers { get; set; } The g2p model’s layers (24).
int G2pMaxNewTokens { get; set; } The g2p model’s most generated tokens (256).
int G2pVocabSize { get; set; } The g2p model’s vocabulary (168,896).
int HopSize { get; set; } Samples per mel frame; paired with the WaveVAE encoder and generator rate products.
int LatentChannels { get; set; } The WaveVAE latent’s channels (32).
int LocalConditionDim { get; set; } The local condition’s width (512).
int NumDecoderLayers { get; set; } The diffusion transformer’s layer count; forwards to DitLayers.
int NumEncoderLayers { get; set; } The phone encoder’s layer count; forwards to PhoneEncoderLayers.
int NumHeads { get; set; } The diffusion transformer’s attention heads; forwards to DitHeads.
int PhoneEncoderHeads { get; set; } The phone encoder’s heads (4).
int PhoneEncoderLayers { get; set; } The phone encoder’s layers (6).
int SamplingSeed { get; set; } The seed of the duration disturbance, the diffusion noise, the posterior draw and language detection.
int Steps { get; set; } Diffusion steps (--time_step, 32).
int TrainingSeed { get; set; } The seed of training’s draws.
int VaeEncoderDimension { get; set; } The encoder’s output width (512).
int VaeEncoderFilters { get; set; } The encoder’s base filters (32).
int VaeGeneratorFilters { get; set; } The generator’s base filters (128).
int VaeLatentUpsampling { get; set; } The latent upsampler’s factor (4).
int VaeUpsamplerChannels { get; set; } The latent upsampler’s channels (320).
int VocabSize { get; set; } The diffusion transformer’s phone vocabulary; forwards to DitPhoneVocabulary.
int WarmupSteps { get; set; } Linear warm-up steps from a rate of 0 (10,000), then the constant rate; 0 trains at the rate from the first step.
int[] VaeEncoderRatios { get; set; } The SEANet encoder’s downsampling ratios (6, 5, 4, 4, 2: 960 samples a latent frame).
int[] VaeGeneratorRatios { get; set; } The generator’s upsampling ratios (5, 4, 4, 3).
string G2pTokenizerJson { get; set; } The g2p tokenizer’s tokenizer.json text; LoadReferenceWeights reads it from the checkpoint.