WaveNet
DeepMind's 2016 neural network that generates audio one sample at a time, and made machine voices noticeably less machine-like.
Earlier speech synthesis either stitched together recorded fragments or drove a parametric vocoder, and sounded like it. WaveNet modelled the waveform itself: each new sample is predicted from all the previous ones, using causal convolutions whose dilation doubles layer by layer so the network can see hundreds of milliseconds into the past. Samples were μ-law quantised to 256 levels, turning the task into classification. In listening tests it closed about half the gap between synthetic and human speech.
The catch was speed — a second of audio originally took minutes to generate. A parallel variant in 2017 solved that and shipped in Google Assistant. The architecture became the template for neural vocoders and turns up in unexpected places, such as amp models for guitarists.