CosyVoiceFamilyOptions

Options & Configuration · AiDotNet.TextToSpeech.CodecBased

Options shared by the CosyVoice models: audio and codec geometry, language-model dimensions, HiFT, CAM++ conditioning, optional released frontend paths and training. Each model sets its release’s defaults.

Inheritance: ModelOptions → CosyVoiceFamilyOptions

Known derived types: CosyVoice3Options, CosyVoiceConformerOptions

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } Exported graph runtime settings.
bool IncludeDiscriminators { get; set; } Whether to build HiFT’s discriminators, which only Vocoder training uses.
double DropoutRate { get; set; } Native encoder dropout probability.
double LearningRate { get; set; } Configured language-model and flow learning rate; the GAN retains its separately declared rate.
double WeightDecay { get; set; } Optional coupled L2 weight decay for native Adam (zero in the released recipes).
int CodebookSize { get; set; } Speech-token vocabulary cardinality.
int CodecFrameRate { get; set; } Speech tokens per second.
int FftSize { get; set; } FFT window size for the conditioning frontend.
int HiftBaseChannels { get; set; } HiFT’s first upsampling width (base_channels).
int HiftHarmonics { get; set; } HiFT’s harmonics above the fundamental (nb_harmonics).
int HopSize { get; set; } Samples per mel frame; must equal four times the product of HiFT upsampling rates.
int LLMDim { get; set; } Language-model hidden width.
int MaxReferenceSeconds { get; set; } Maximum prompt duration accepted by the released frontend (30 seconds).
int MaxTextLength { get; set; } Maximum input text length.
int MelChannels { get; set; } Mel-spectrogram frequency channels.
int NumCodebooks { get; set; } The single S3 token stream used by the CosyVoice family.
int NumHeads { get; set; } Language-model query attention heads.
int NumLLMLayers { get; set; } Language-model decoder blocks.
int SampleRate { get; set; } Audio sample rate in Hz.
int SamplingSeed { get; set; } The seed of sampling and of training’s random draws.
int SpeakerEmbeddingDim { get; set; } Flow conditioning x-vector width.
int SpeakerEncoderChannels { get; set; } CAM++’s TDNN channels (init_channels).
int SpeakerEncoderGrowthRate { get; set; } CAM++’s D-TDNN growth rate.
int VocabSize { get; set; } Text vocabulary size.
int VocoderSegmentLength { get; set; } The samples a vocoder training step crops or pads an utterance to (the recipe’s truncate_length); a multiple of HopSize.
int WarmupSteps { get; set; } The learning-rate warmup of each stage’s Adam (the recipe’s warmup_steps; 0 for a constant rate).
int[] HiftSourceKernels { get; set; } HiFT’s source-branch residual block kernels, one per upsampling stage.
int[] HiftUpsampleKernels { get; set; } HiFT’s transposed-convolution kernels, one per upsampling stage.
int[] HiftUpsampleRates { get; set; } HiFT’s upsampling rates; their product times the iSTFT hop of 4 is HopSize.
string SpeakerEncoderOnnxPath { get; set; } Optional original frozen CAM++ graph; release loading sets this path.
string SpeechTokenizerPath { get; set; } Optional original frozen speech-tokenizer ONNX graph; release loading sets this path.