MetaVoice1BOptions
Options & Configuration · AiDotNet.TextToSpeech.VoiceCloning
Options for the MetaVoice-1B multi-stage voice-cloning TTS model.
Inheritance: ModelOptions → TtsModelOptions → CodecTtsOptions → VoiceCloningOptions → MetaVoice1BOptions
For Beginners
These options configure the MetaVoice-1B model. The defaults reproduce the released first-stage causal GPT, second-stage noncausal GPT, speaker encoder, eight-codebook EnCodec and four-band diffusion decoder.
Constructors
| Constructor | Summary |
|---|---|
MetaVoice1BOptions() |
Initializes a new instance with default values. |
MetaVoice1BOptions(MetaVoice1BOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool EnhanceOutput { get; set; } |
Runs the release’s DeepFilterNet3 stage and returns 48 kHz audio. |
bool RawDecoderOutput { get; set; } |
Returns raw 24 kHz neural output when true; the default reproduces the release’s loudness and PCM policy. |
double FirstStageTemperature { get; set; } |
First-stage logit temperature. |
double PromptGuidance { get; set; } |
Text-prompt classifier-free guidance scale. |
double RoPETheta { get; set; } |
Legacy factory setting; the released native transformers use learned absolute positions. |
double SecondStageTemperature { get; set; } |
Second-stage logit temperature. |
double SpeakerGuidance { get; set; } |
Speaker classifier-free guidance scale. |
double TrainingDropout { get; set; } |
Original fine-tuning embedding, attention and residual dropout. |
double? FirstStageTopP { get; set; } |
First-stage nucleus probability. |
int CodecLatentDim { get; set; } |
EnCodec continuous-latent width (summed RVQ codebook embedding dim); the vocoder’s input channel count. |
int DecoderDim { get; set; } |
Legacy decoder width (retained for serialization compatibility). |
int EncoderDim { get; set; } |
Legacy encoder width (retained for serialization compatibility). |
int FineTuneLastBlocks { get; set; } |
Original fine-tuning freezes everything except the final norm and these last blocks; zero selects all blocks. |
int FirstStageCodebooks { get; set; } |
EnCodec RVQ hierarchies predicted by the first stage (paper: the first two). |
int FirstStageContext { get; set; } |
First-stage learned-position context. |
int FirstStageDim { get; set; } |
First-stage causal GPT width, divisible by the first-stage head count. |
int MaximumNewTokens { get; set; } |
Original canonical generation budget. |
int NumFirstStageLayers { get; set; } |
Number of first-stage causal transformer blocks. |
int NumSecondStageLayers { get; set; } |
Number of second-stage non-causal transformer blocks. |
int ReferenceSampleRate { get; set; } |
Sample rate of tensor-only reference recordings, before resampling to 16 kHz. |
int SecondStageContext { get; set; } |
Second-stage fixed padded context. |
int SecondStageDim { get; set; } |
Released noncausal GPT width. |
int SecondStageHeads { get; set; } |
Released noncausal GPT head count. |
int SpeakerEncoderLayers { get; set; } |
Released speaker LSTM depth. |
int SpeakerHiddenDim { get; set; } |
Released speaker LSTM width. |
int SpeakerMelChannels { get; set; } |
Released speaker power-mel channels. |
int SwiGLUMultipleOf { get; set; } |
Rounds SwiGLU FFN hidden dimensions to this multiple. |
int VocoderChannels { get; set; } |
Legacy factory setting retained for serialized options; the native decoder is multiband diffusion. |
int VocoderUpsampleFactor { get; set; } |
Legacy factory setting retained for serialized options. |
int? FirstStageTopK { get; set; } |
Optional first-stage top-k; choose either this or top-p. |
int? SecondStageTopK { get; set; } |
Second-stage top-k filtering. |
int? SynthesisSeed { get; set; } |
Optional seed replayed for each synthesis. |