Parallel / non-autoregressive

FastSpeech threw away left-to-right spectrogram decoding. A duration model expands phonemes; a feed-forward Transformer paints the mel in parallel. Inference became a vocoder-bound problem again, and skipped/repeated words became rarer.

FastSpeech 2 trained duration, pitch, and energy against real targets instead of a Tacotron teacher. That made the system controllable: you can raise pitch without retraining. A lot of later ‘fast’ commercial stacks are this skeleton with a better vocoder.

The trade is oversmoothing. Parallel models average possible readings. Later GAN vocoders, then VITS, then flow-matching systems, are attempts to put the missing variance back without returning to slow autoregression.