VallEModelOptions

Options & Configuration ยท AiDotNet.TextToSpeech.CodecBased

The native codec, Transformer and decoding configuration shared by the VALL-E family.

Inheritance: ModelOptions โ†’ VallEModelOptions

Known derived types: VALLE2Options, VALLEOptions, VALLEXOptions

How It Works

Concrete configurations add their own prompt policies: VALL-E samples a duration, VALL-E X reads an explicit second utterance, and VALL-E 2 samples its own duration range.

Example

using AiDotNet.TextToSpeech.CodecBased;
var options = new VALLEXOptions { HiddenDim = 1024, NumHeads = 16 };

Properties

Property Summary
EnCodecOptions Codec { get; set; } The pretrained codec (EnCodec 24 kHz at 6 kbps by default).
OnnxModelOptions OnnxOptions { get; set; } Settings for the exported ONNX graph runtime.
bool PrependBos { get; set; } Prepend the beginning-of-audio token used by the community VALL-E-X checkpoint.
bool StopOnMostLikelyEndToken { get; set; } Stop when EOS is most likely, even if sampling selected another token.
double DropoutRate { get; set; } The Transformer training dropout.
double LearningRate { get; set; } The peak AdamW learning rate.
double Temperature { get; set; } The AR sampling temperature.
double WeightDecay { get; set; } The AdamW weight decay.
int CodebookSize { get; set; } The codes per codebook, matching the codec.
int FeedForwardDim { get; set; } The feed-forward width of both Transformers.
int HiddenDim { get; set; } The hidden width of both Transformers.
int HopSize { get; set; } The samples per codec frame, matching the codec strides.
int MaxCodesPerTextToken { get; set; } The AR generation limit in codes per phoneme.
int MaxTextLength { get; set; } The phonemes retained by the frontend before framing.
int NumCodebooks { get; set; } The codec codebooks modeled by AR and NAR together.
int NumDecoderLayers { get; set; } The AR Transformer depth, equal to the NAR depth.
int NumEncoderLayers { get; set; } The NAR Transformer depth, equal to the AR depth.
int NumHeads { get; set; } The attention heads of both Transformers.
int SampleRate { get; set; } The waveform sample rate, matching the codec.
int SamplingSeed { get; set; } The seed of training draws and sampling.
int TextTokens { get; set; } The phoneme embedding rows, sufficient for the selected inventory.
int TopK { get; set; } The AR top-k cut, or zero to retain every code.
int TrainingSteps { get; set; } The updates over which the learning rate decays linearly to zero.
int WarmupSteps { get; set; } The linear warmup duration in updates.
string ModelPath { get; set; } The exported ONNX graph path, when a graph is loaded.