fastspeech

Definition

A non-autoregressive TTS model that produces all acoustic frames in parallel instead of one at a time. A duration predictor and length regulator expand the phoneme sequence to the right number of frames, which removes the sequential dependency that makes autoregressive synthesis slow. The design also avoids the word-skipping and repetition failures that attention-based autoregressive TTS is prone to, since alignment is predicted explicitly rather than learned implicitly.