VITSOptions

Options & Configuration · AiDotNet.TextToSpeech.EndToEnd

Options for VITS (Kim et al. 2021): a conditional VAE with a normalizing-flow prior, a stochastic duration predictor and a HiFi-GAN decoder trained adversarially end to end.

Inheritance: ModelOptions → TtsModelOptions → EndToEndTtsOptions → VitsModelOptions → VITSOptions

Known derived types: PiperOptions, YourTTSOptions

For Beginners

These options configure the VITS model. Default values follow the original paper settings.

How It Works

The shared settings are in VitsModelOptions. VITS adds its stochastic duration predictor (kernel 3, dropout 0.5, 4 spline flows; the reference’s StochasticDurationPredictor(hidden, 192, 3, 0.5, 4)) sampled with noise scale 0.8 at synthesis, and places blank tokens between the characters.

Constructors

Constructor Summary
VITSOptions() Creates the paper’s configuration.
VITSOptions(VITSOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
double DurationNoiseScale { get; set; } Gets or sets the duration-predictor noise scale at synthesis (0.8).
int DurationPredictorFlows { get; set; } Gets or sets the stochastic duration predictor’s spline flows (4).