WaveNet

Text-to-speech in 2016 either stitched together fragments of recorded speech or drove a vocoder with a statistical model, and both sounded like machines. WaveNet, announced by DeepMind on 8 September 2016, modelled the audio waveform directly, predicting each of the 16,000 samples per second from the ones before it, with a stack of convolutions whose dilations doubled at each layer so that the network could see a quarter of a second of context. Listeners rated its output halfway between the previous best system and a human recording; nothing before had closed that gap by half.

It was absurdly slow at first, generating a second of audio in several minutes, and its arrival in Google Assistant in 2017 required a redesigned parallel version. But the demonstration mattered beyond speech. WaveNet showed that autoregressive prediction, the next-sample or next-word objective, worked on raw signals with no hand-built representation in between, and the same dilated-convolution idea reappeared in models of text and time series.

The voices that read this site's events, the assistants that talk back, and the cloned voices that regulators worry about all descend from it. Synthetic speech that people cannot distinguish from a recording dates from this paper.