IVocoder<T>

Interfaces · AiDotNet.TextToSpeech.Interfaces

Interface for neural vocoders that convert mel-spectrograms to audio waveforms.

Known implementations: VocoderBase<T>

How It Works

Vocoders form the second stage of a two-stage TTS pipeline: Text -> [Acoustic Model] -> Mel-Spectrogram -> [Vocoder] -> Waveform. Architectures include:

  • Autoregressive: WaveNet, WaveRNN (sample-by-sample generation)
  • GAN-based: HiFi-GAN, MelGAN, BigVGAN (adversarial training, parallel)
  • Flow-based: WaveGlow (invertible flow, parallel)
  • Diffusion-based: DiffWave, WaveGrad (denoising diffusion)
  • Fourier-based: Vocos, iSTFTNet (inverse STFT output)

Properties

Property Summary
int MelChannels { get; } Gets the expected number of mel channels in the input spectrogram.
int SampleRate { get; } Gets the audio sample rate in Hz (e.g., 22050, 24000).
int UpsampleFactor { get; } Gets the upsampling factor (hop_size) from mel frames to audio samples.

Methods

Method Summary
Tensor<T> MelToWaveform(Tensor<T> melSpectrogram) Converts a mel-spectrogram to an audio waveform.