wavenet

Definition

A DeepMind network that generates raw audio one sample at a time, using stacked dilated causal convolutions to reach far back in time without an enormous number of layers. It showed that neural waveform generation could beat concatenative and parametric synthesis on naturalness, which reset expectations for TTS. Its weakness is speed — autoregressive sampling at audio rates is slow — and much later work exists to remove exactly that bottleneck.

Explainers using this term

• Speech Synthesis from Scratch — From Text to a VoiceWaveNet: A Generative Model for Raw Audio