MultiBandMelGANOptions

Options & Configuration · AiDotNet.TextToSpeech.Vocoders

Options for Multi-band MelGAN (Yang et al. 2021): a MelGAN generator predicting four PQMF sub-bands, trained with the full- and sub-band multi-resolution STFT loss after a generator-only pre-training phase.

Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → MultiBandMelGANOptions

For Beginners

These options configure the Multi-band MelGAN model. Default values follow the original paper settings.

How It Works

Defaults are the paper’s (§3.1–3.2, Tables 1–2): 16 kHz audio, 50 ms frames (800 samples) with a 12.5 ms shift (200) and a 1024-point FFT, 80 mel bands standardized per band; a generator of 384 first channels upsampling ×2, ×5, ×5 to 192, 96, 48 channels, residual stacks of four blocks (dilations 1, 3, 9, 27) and four output bands merged by a PQMF of 63 coefficients; three discriminators with strided convolutions of 64, 256, 512 channels; LSGAN; the full-band STFT resolutions (1024, 600, 120), (2048, 1200, 240), (512, 240, 50) and sub-band ones (384, 150, 30), (683, 300, 60), (171, 60, 10); 200k generator-only steps; Adam at 1e-4 for both networks, halved every 100k steps down to 1e-6; one second of audio per example, batch 128.

What the paper leaves open follows kan-bayashi/ParallelWaveGAN (multi_band_melgan.v2.yaml, logmelfilterbank): λ_adv = 2.5 on the adversarial term; mel bands from 80 to 7600 Hz with log10(max(x, 1e-10)); PQMF cutoff 0.142 and Kaiser β 9.

Constructors

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

Properties

Property Summary
double AdversarialWeight { get; set; } Gets or sets λ_adv, the adversarial term’s weight (2.5).
double MelMaxFrequency { get; set; } Gets or sets the highest mel frequency (7600 Hz).
double MelMinFrequency { get; set; } Gets or sets the lowest mel frequency (80 Hz).
double MinimumLearningRate { get; set; } Gets or sets the learning rate’s floor (1e-6).
double PqmfBeta { get; set; } Gets or sets the PQMF prototype’s Kaiser β (9).
double PqmfCutoffRatio { get; set; } Gets or sets the PQMF prototype’s cutoff ratio (0.142).
double[] MelMean { get; set; } Gets or sets the per-band mean of the log-mel features over the training set, or null for none.
double[] MelScale { get; set; } Gets or sets the per-band standard deviation of the log-mel features over the training set, or null for none.
int DiscriminatorWidthDivisor { get; set; } Gets or sets a divisor on every discriminator width (1: the paper’s widths).
int LearningRateHalvingSteps { get; set; } Gets or sets the steps between learning-rate halvings (100,000).
int NumBands { get; set; } Gets or sets the number of sub-bands (4).
int NumDiscriminators { get; set; } Gets or sets the number of discriminators (3).
int PqmfTaps { get; set; } Gets or sets the PQMF taps (62: filters of 63 coefficients).
int ResidualLayers { get; set; } Gets or sets the residual blocks per stack (4: dilations 1, 3, 9, 27).
int SamplingSeed { get; set; } Gets or sets the seed of the training segment draws.
int SegmentSize { get; set; } Gets or sets the training segment in samples (16,000: one second).
int WindowSize { get; set; } Gets or sets the analysis window of the input features (800: 50 ms).
int[] DiscriminatorChannels { get; set; } Gets or sets the discriminators’ strided convolution widths (64, 256, 512).
int[] FullBandFftSizes { get; set; } Gets or sets the full-band STFT loss FFT sizes (1024, 2048, 512).
int[] FullBandHopSizes { get; set; } Gets or sets the full-band STFT loss hops (120, 240, 50).
int[] FullBandWindowSizes { get; set; } Gets or sets the full-band STFT loss windows (600, 1200, 240).
int[] SubBandFftSizes { get; set; } Gets or sets the sub-band STFT loss FFT sizes (384, 683, 171).
int[] SubBandHopSizes { get; set; } Gets or sets the sub-band STFT loss hops (30, 60, 10).
int[] SubBandWindowSizes { get; set; } Gets or sets the sub-band STFT loss windows (150, 300, 60).
long PretrainingSteps { get; set; } Gets or sets the generator-only pre-training steps (200,000).