BigVGAN<T>

Models & Types · AiDotNet.TextToSpeech.Vocoders

BigVGAN: a universal GAN vocoder — HiFi-GAN with anti-aliased periodic (Snake) activations, trained at scale against multi-period and multi-resolution discriminators.

Inheritance: NeuralNetworkBase<T> → TtsModelBase<T> → VocoderBase<T> → GanVocoderBase<T> → BigVGAN<T>

For Beginners

BigVGAN turns spectrograms into audio for any voice, language or instrument; its activations ripple periodically like sound does, and a filter keeps them from adding high-frequency artefacts.

How It Works

References: “BigVGAN: A Universal Neural Vocoder with Large-Scale Training” (Lee et al., ICLR 2023) and NVIDIA/BigVGAN for what the paper leaves unstated.

The generator replaces HiFi-GAN’s leaky ReLUs with Snake, made anti-aliased by 2× sinc up- and down-sampling around it (the AMP module, §3.2–3.3). Training is HiFi-GAN’s with the multi-scale discriminator replaced by UnivNet’s multi-resolution discriminator (§3.1): LSGAN, feature matching (×2) and mel L1 (×45), AdamW (0.8, 0.99) at 1e-4 decayed 0.999 per epoch, gradient norms clipped at 1000 — the generator’s, the period discriminators’ and the resolution discriminators’ separately (§3.4, reference train.py).

Constructors

Constructor Summary
BigVGAN(NeuralNetworkArchitecture<T> architecture, BigVGANOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a trainable BigVGAN.
BigVGAN(NeuralNetworkArchitecture<T> architecture, string modelPath, BigVGANOptions options) Creates a trainable BigVGAN.

Properties

Property Summary
int UpsampleFactor { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()