EmotiVoiceOptions

Options & Configuration · AiDotNet.TextToSpeech.StyleEmotion

Options for EmotiVoice.

Inheritance: ModelOptions → TtsModelOptions → EmotiVoiceOptions

For Beginners

These settings size EmotiVoice’s networks; the defaults reproduce the released model.

How It Works

EmotiVoice has no paper: the defaults are its released configuration (netease-youdao/EmotiVoice config/joint/config.py and config.yaml): 16 kHz audio, 80 mel bins up to 8 kHz from a 1024-point FFT with a 256-sample hop; 384-wide four-block Transformer encoder and decoder (8 heads, kernel-3 feed-forward of 1536, dropout 0.2); 2014 speakers and 502 phoneme tokens; SimBERT (BERT-base Chinese, 13,685 tokens) embeddings of the emotion prompt and of the text; HiFi-GAN V1 (upsampling 8·8·2·2) trained jointly on 32-frame segments; Adam (rate 1.25e-5, β = 0.5, 0.9) with an exponential decay of 0.999875 per epoch.

Constructors

Constructor Summary
EmotiVoiceOptions() Creates the released configuration.
EmotiVoiceOptions(EmotiVoiceOptions other) Copies another instance.

Properties

Property Summary
bool TrimSilence { get; set; } Whether training trims the recording’s silence (below 0.5 % of its peak) and pads 50 ms on each side, as the released data loader does (true).
double Beta1 { get; set; } Adam β₁ (0.5).
double Beta2 { get; set; } Adam β₂ (0.9).
double BinarizationWeight { get; set; } Weight of the binarization loss (2).
double DurationDropout { get; set; } Dropout of the duration predictor (0.5).
double EnergyMean { get; set; } Energy normalization mean (30.610).
double EnergyStd { get; set; } Energy normalization standard deviation (21.78).
double Epsilon { get; set; } Adam ε (1e-9).
double ForwardSumWeight { get; set; } Weight of the forward-sum loss (2).
double LearningRateDecay { get; set; } Learning-rate decay per epoch (0.999875).
double MelLossWeight { get; set; } Weight of the mel L1 loss (45).
double MelMaxFrequency { get; set; } Highest mel frequency (8000 Hz).
double MelMinFrequency { get; set; } Lowest mel frequency (0 Hz).
double PitchMean { get; set; } Pitch normalization mean (225.089 Hz).
double PitchStd { get; set; } Pitch normalization standard deviation (53.78).
double VarianceDropout { get; set; } Dropout of the pitch and energy predictors (0.1).
int DecoderLayers { get; set; } Decoder blocks (4).
int DiscriminatorWidthDivisor { get; set; } A divisor on every width of HiFi-GAN’s multi-period and multi-scale discriminators (1: the released 32–1024 channels).
int DurationKernel { get; set; } Kernel of the duration predictor (3).
int DurationLayers { get; set; } Convolutions of the duration predictor (2).
int EmotionLabels { get; set; } Emotion classes of the style encoder’s pre-training (7: 普通 生气 开心 惊讶 悲伤 厌恶 恐惧).
int EncoderLayers { get; set; } Encoder blocks (4).
int EnergyLabels { get; set; } Energy classes (3).
int EnergyLayers { get; set; } Convolutions of the energy predictor (2).
int FeedForwardKernel { get; set; } Kernel of the blocks’ convolutional feed-forward (3); its width is 4 × HiddenDim.
int Heads { get; set; } Attention heads of the encoder and decoder (8).
int PitchLabels { get; set; } Pitch classes (3: normal, high, low).
int PitchLayers { get; set; } Convolutions of the pitch predictor (3).
int SamplingSeed { get; set; } Seed of the training draws and of dropout.
int SegmentFrames { get; set; } Frames of the random segment the vocoder trains on (32).
int SpeakerCount { get; set; } Speakers in the speaker table (2014, the released list).
int SpeedLabels { get; set; } Speed classes (3).
int StyleBertHeads { get; set; } SimBERT’s heads (12).
int StyleBertHidden { get; set; } SimBERT’s width (768).
int StyleBertIntermediate { get; set; } SimBERT’s feed-forward width (3072).
int StyleBertLayers { get; set; } SimBERT’s layers (12).
int UpdatesPerEpoch { get; set; } Updates per epoch, after which the learning rate decays (0: never; the released code decays once per epoch of your data — set it for yours).
int UpsampleInitialChannels { get; set; } HiFi-GAN’s first width (512).
int VarianceChannels { get; set; } Channels of the variance predictors (384).
int VarianceEmbedKernel { get; set; } Kernel of the pitch and energy embedding convolutions (9).
int VarianceKernel { get; set; } Kernel of the pitch and energy predictors (3).
int WindowLength { get; set; } The analysis window, in samples (1024).
int[] ResBlockKernels { get; set; } HiFi-GAN’s residual-block kernels (3, 7, 11).
int[] UpsampleKernels { get; set; } HiFi-GAN’s upsampling kernels (16, 16, 4, 4).
int[] UpsampleRates { get; set; } HiFi-GAN’s upsampling factors (8, 8, 2, 2); their product is the hop.
int[][] ResBlockDilations { get; set; } Their dilations ({1, 3, 5} each).