CosyVoice3Options

Options & Configuration · AiDotNet.TextToSpeech.CodecBased

Options for the Fun-CosyVoice3-0.5B-2512 release.

Inheritance: ModelOptions → CosyVoiceFamilyOptions → CosyVoice3Options

How It Works

The defaults reproduce the released Qwen2 LM, causal DiT flow and 24 kHz CausalHiFT.

Constructors

Constructor Summary
CosyVoice3Options() Initializes a new instance with default values.
CosyVoice3Options(CosyVoice3Options other) Initializes a copy.

Properties

Property Summary
double DiTInferenceCfgRate { get; set; } Gets or sets the diffusion transformer inference classifier-free guidance rate.
double DiTSigmaMin { get; set; } Gets or sets the diffusion transformer sigma min.
double DiTTrainingCfgRate { get; set; } Gets or sets the diffusion transformer training classifier-free guidance rate.
double RepetitionRatio { get; set; } Repetition-aware LM sampler threshold ratio.
double RmsEpsilon { get; set; } Qwen2 RMS normalization epsilon.
double RopeTheta { get; set; } Qwen2 rotary base.
double TopP { get; set; } Repetition-aware LM sampler nucleus probability.
int DiTDepth { get; set; } Gets or sets the diffusion transformer depth.
int DiTDim { get; set; } Gets or sets the feature width of the diffusion transformer.
int DiTFeedForwardMultiplier { get; set; } Gets or sets the diffusion transformer feed forward multiplier.
int DiTHeadDim { get; set; } Gets or sets the feature width of the diffusion transformer head.
int DiTHeads { get; set; } Gets or sets the number of attention heads in the diffusion transformer.
int DiTInputDim { get; set; } Gets or sets the feature width of the diffusion transformer input.
int DiTTimesteps { get; set; } Gets or sets the diffusion transformer timesteps.
int KeyValueHeads { get; set; } Qwen2 key-value attention heads.
int LmFeedForwardDim { get; set; } Qwen2 feed-forward width.
int Lookahead { get; set; } Speech tokens read ahead by the flow’s pre-lookahead convolution.
int PreLookaheadChannels { get; set; } Gets or sets the pre lookahead channels.
int RepetitionWindow { get; set; } Repetition-aware LM sampler history window.
int ReservedSpeechTokens { get; set; } Reserved speech ids after the S3 vocabulary; every one is a stop id during generation.
int StaticChunkSize { get; set; } Streaming DiT chunk size in speech tokens.
int TokenMelRatio { get; set; } The released DiT flow produces exactly two mel frames per speech token.
int TopK { get; set; } Repetition-aware LM sampler candidate count.