APNet2Options
Options & Configuration · AiDotNet.TextToSpeech.Vocoders
Options for APNet2 (Du et al. 2023): APNet with ConvNeXt v2 amplitude and phase predictors, linear anti-wrapping phase losses, and a hinge GAN loss against a multi-period and a multi-resolution discriminator.
Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → AmplitudePhaseVocoderOptions → APNet2Options
For Beginners
These options configure the APNet2 model. Default values follow the original paper settings.
How It Works
Defaults are the paper’s (§3–4.1): 22.05 kHz audio; 80-band mel spectrograms (0–8 kHz) and spectra from a 1024-point FFT with a 1024-sample window and a 256-sample shift; 8 ConvNeXt v2 blocks of 512 channels with depth-wise kernel 7 and a 1536-wide point-wise expansion in both predictors; APNet’s loss weights; AdamW (0.8, 0.99, weight decay 0.01) at 2e-4 decayed by 0.999 per epoch; batch 16 and 8,192-sample segments.
What the paper leaves open follows redmist328/APNet2: input and output convolutions with kernel 7 and every convolution and linear map initialized from N(0, 0.02²) truncated at ±2 with zero biases; the MRD of three resolutions (FFT, hop, window) = (1024, 256, 1024), (2048, 512, 2048), (512, 128, 512) over unwindowed magnitude spectrograms with 64-channel strided 2-D convolutions and leaky ReLU 0.1; the feature-matching weight 1, 0.1 on the MRD’s feature maps; 750 updates per epoch (12,000 clips at batch 16). The hinge GAN losses average the MPD’s and MRD’s sub-discriminators equally, as the paper’s Eq. 2–3 state (the reference sums them and weights the MRD’s by 0.1).
Constructors
| Constructor | Summary |
|---|---|
APNet2Options() |
Creates the paper’s APNet2 configuration. |
APNet2Options(APNet2Options other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double FeatureMatchingWeight { get; set; } |
Gets or sets the feature-matching weight (1). |
double ResolutionFeatureMatchingWeight { get; set; } |
Gets or sets the weight of the MRD’s feature-matching terms relative to the MPD’s (0.1). |
int ConvNeXtChannels { get; set; } |
Gets or sets the ConvNeXt v2 channels (512). |
int ConvNeXtIntermediateChannels { get; set; } |
Gets or sets the ConvNeXt v2 point-wise expansion (1536). |
int DepthwiseKernelSize { get; set; } |
Gets or sets the depth-wise convolution’s kernel (7). |
int NumConvNeXtBlocks { get; set; } |
Gets or sets the ConvNeXt v2 blocks per predictor (8). |
int ResolutionDiscriminatorChannels { get; set; } |
Gets or sets the MRD’s channels (64). |
int[] ResolutionFftSizes { get; set; } |
Gets or sets the MRD’s FFT sizes (1024, 2048, 512). |
int[] ResolutionHopSizes { get; set; } |
Gets or sets the MRD’s hops (256, 512, 128). |
int[] ResolutionWindowSizes { get; set; } |
Gets or sets the MRD’s windows (1024, 2048, 512). |