APNetOptions
Options & Configuration · AiDotNet.TextToSpeech.Vocoders
Options for APNet (Ai and Ling 2023): an all-frame-level vocoder that predicts the log amplitude and the phase spectra directly and reconstructs the waveform by an inverse STFT.
Inheritance: ModelOptions → TtsModelOptions → VocoderOptions → AmplitudePhaseVocoderOptions → APNetOptions
For Beginners
These options configure the APNet model. Default values follow the original paper settings.
How It Works
Defaults are the paper’s (§IV): 16 kHz audio; 80-band mel spectrograms and spectra from a 1024-point FFT with a 320-sample (20 ms) window and an 80-sample (5 ms) shift; ASP and PSP residual networks of 3 parallel blocks with kernels 3, 7, 11 and 3 subblocks with dilations 1, 3, 5, 512 channels; the shared loss weights; HiFi-GAN’s MPD and MSD with the least-squares GAN loss; batch 16 and 8,000-sample segments.
What the paper leaves open follows yangai520/APNet: input and output convolutions with kernel 7; mel spectrograms
from a centred STFT, Slaney bands 0–8 kHz, ln(max(x, 1e-5)); the feature-matching weight 2; 655 updates per
epoch (10,480 training utterances at batch 16).
Constructors
| Constructor | Summary |
|---|---|
APNetOptions() |
Creates the paper’s APNet configuration. |
APNetOptions(APNetOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
double FeatureMatchingWeight { get; set; } |
Gets or sets the feature-matching weight (2). |
int Channels { get; set; } |
Gets or sets the channels of the ASP and PSP networks (512). |
int[][] ResblockDilationSizes { get; set; } |
Gets or sets the dilations of each block’s subblocks (1, 3, 5). |