VevoOptions

Options & Configuration · AiDotNet.TextToSpeech.CodecBased

Options for Vevo.

Inheritance: ModelOptions → TtsModelOptions → VevoOptions

For Beginners

These settings size Vevo’s networks and control how it imitates a voice; the defaults reproduce the released model.

How It Works

The defaults are the released Vevo (Zhang et al., ICLR 2025; Amphion models/vc/vevo and the amphion/Vevo checkpoints): HuBERT-Large read at its 18th layer; a content-style tokenizer of 8,192 codes (RepCodec over the normalized features) and a content tokenizer of 32 (the original RepCodec over the raw ones); a 1920-wide, 12-layer Llama from phones to content-style tokens (Vevo-TTS) and a 1536-wide one from content tokens (Vevo-Voice); a 1024-wide, 16-layer flow-matching transformer to a 128-band mel at 24 kHz, hop 480; and a 30-block Vocos.

Decoding samples at temperature 0.8, top-k 50, top-p 0.9, with at least 50 new tokens; the flow takes 32 Euler steps with guidance 1 rescaled by 0.75. Training follows the release’s trainers: AdamW at 2e-4 for the autoregressive models (gradient norm clipped at 1) and 7.5e-5 for the flow (clipped at 0.2), an inverse-square-root schedule after 32,000 warm-up steps.

Constructors

Constructor Summary
VevoOptions() Creates the released Vevo configuration.

Properties

Property Summary
VevoStage CurrentStage { get; set; } The network Vevo’s training updates.
bool ConstrainToContentStyleTokens { get; set; } Restrict decoding to content-style tokens and the end token.
bool Sample { get; set; } Sample the next token (true, the release); otherwise take the most likely.
bool VoiceConversion { get; set; } Whether Vevo-Voice and Vevo-Style (the content tokenizer and the second transformer) are built.
double Beta1 { get; set; } AdamW’s β1 (0.9).
double Beta2 { get; set; } AdamW’s β2 (0.999).
double Epsilon { get; set; } AdamW’s ε (1e-8).
double FlowLearningRate { get; set; } The flow’s learning rate (7.5e-5); the autoregressive models use LearningRate.
double FlowPromptDropProbability { get; set; } How often training drops the prompt and the condition (cfg_scale, 0.2).
double FlowSigma { get; set; } σ of the probability path (1e-5).
double Guidance { get; set; } Guidance scale (1).
double GuidanceRescale { get; set; } The share of the rescaled guided velocity (0.75).
double MelMaxFrequency { get; set; } The mel’s upper edge (12 kHz).
double MelMean { get; set; } The mel’s normalization mean (−4.92).
double MelVariance { get; set; } The mel’s normalization variance (8.14).
double RepetitionPenalty { get; set; } Repetition penalty (1, none).
double Temperature { get; set; } Temperature (0.8).
double TopP { get; set; } Top-p (0.9).
int ContentCodecChannels { get; set; } The content tokenizer’s width and code dimension (1024).
int ContentCodes { get; set; } Content tokens (32).
int ContentSampleRate { get; set; } The 16 kHz rate HuBERT reads.
int ContentStyleCodeDim { get; set; } The content-style codebook’s dimension (8).
int ContentStyleCodes { get; set; } Content-style tokens (8192).
int ContentStyleVocosDim { get; set; } The content-style tokenizer’s Vocos width (384).
int ContentStyleVocosIntermediate { get; set; } The content-style tokenizer’s Vocos feed-forward width (2048).
int ContentStyleVocosLayers { get; set; } The content-style tokenizer’s Vocos blocks (12).
int FlowHeads { get; set; } The flow-matching transformer’s heads (16).
int FlowLayers { get; set; } The flow-matching transformer’s layers (16).
int FlowSteps { get; set; } Euler steps (32).
int HubertHeads { get; set; } HuBERT’s heads (16).
int HubertHidden { get; set; } HuBERT’s width (1024).
int HubertIntermediate { get; set; } HuBERT’s feed-forward width (4096).
int HubertLayer { get; set; } The HuBERT layer whose output the tokenizers read (18); the layers above it are never run.
int HubertPositionGroups { get; set; } HuBERT’s position-convolution groups (16).
int HubertPositionKernel { get; set; } HuBERT’s position-convolution kernel (128).
int MaxLength { get; set; } The most positions a sequence reaches, input included (2000).
int MinNewTokens { get; set; } New tokens before the end token is allowed (50).
int NumDecoderLayers { get; set; } Gets or sets Vevo-TTS’s actual autoregressive decoder depth.
int NumEncoderLayers { get; set; } Gets or sets the frozen HuBERT encoder depth at which features are read.
int NumHeads { get; set; } Gets or sets Vevo-TTS’s actual autoregressive attention heads.
int PhoneVocabulary { get; set; } Phone vocabulary of the autoregressive transformers (1024).
int SamplingSeed { get; set; } The seed of decoding’s sampling and the flow’s noise.
int TextArHeads { get; set; } Vevo-TTS’s transformer heads (16).
int TextArHidden { get; set; } Vevo-TTS’s transformer width (1920).
int TextArIntermediate { get; set; } Vevo-TTS’s transformer feed-forward width (7680).
int TextArLayers { get; set; } Vevo-TTS’s transformer layers (12).
int TopK { get; set; } Top-k (50).
int TrainingSeed { get; set; } The seed of training’s draws.
int VocoderDim { get; set; } Vocos’s width (1024).
int VocoderIntermediate { get; set; } Vocos’s feed-forward width (4096).
int VocoderLayers { get; set; } Vocos’s blocks (30).
int VoiceArHeads { get; set; } Vevo-Voice’s transformer heads (16).
int VoiceArHidden { get; set; } Vevo-Voice’s transformer width (1536).
int VoiceArIntermediate { get; set; } Vevo-Voice’s transformer feed-forward width (6144).
int VoiceArLayers { get; set; } Vevo-Voice’s transformer layers (12).
int WarmupSteps { get; set; } Warm-up steps of the inverse-square-root schedule (32,000).
int[] HubertConvDims { get; set; } HuBERT’s feature-encoder widths (HuBERT-Large: seven of 512).
int[] HubertConvKernels { get; set; } HuBERT’s feature-encoder kernels.
int[] HubertConvStrides { get; set; } HuBERT’s feature-encoder strides (×320).
string Language { get; set; } The language of the text to speak (g2p_’s: zh, en, ja, ko, fr, de).
string OpenJTalkDictionaryPath { get; set; } The Open JTalk dictionary (open_jtalk_dic_utf_8-1.11), which Japanese text needs.
string PolyphoneModelPath { get; set; } MaskGCT’s polyphone model (poly_bert_model.onnx), which Chinese text needs.