Piper<T>

Models & Types · AiDotNet.TextToSpeech.EndToEnd

Piper: a fast, local neural text-to-speech system — VITS trained with Rhasspy’s recipe and model sizes.

Inheritance: NeuralNetworkBase<T> → TtsModelBase<T> → VitsTtsModelBase<T> → StochasticDurationVitsBase<T> → Piper<T>

For Beginners

Piper is a small, fast VITS voice meant to run on ordinary computers and devices; it turns phonemes straight into audio.

How It Works

References: rhasspy/piper (piper_train, 2023), which trains VITS (Kim et al. 2021) unchanged; see PiperOptions for the sizes and framing it sets.

The network, its training and its synthesis are VITS’s (VITS, with the same stochastic duration predictor). Piper reads phoneme ids framed as piper-phonemize frames them — beginning-of-sentence, then every id followed by the padding id, then end-of-sentence (^ _ p₁ _ p₂ _ … pₙ _ $) — and its medium and x-low voices decode with a lighter HiFi-GAN.

Constructors

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

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()