WaveNet

A deep neural network for generating raw audio waveforms that bridges the gap between machine speech and human performance.

Developed by Google DeepMind in 2016, WaveNet utilizes a dilated causal convolutional architecture to model raw audio at 16,000 samples per second. Unlike traditional concatenative systems that stitch together recorded fragments, this generative model predicts the probability distribution of the next individual sample based on all previous steps. It reduced the gap between human speech and text-to-speech (TTS) quality by over 50 percent on Mean Opinion Scores (MOS) for both English and Mandarin. Today, the technology powers Google Assistant and various cloud-based voice synthesis applications, delivering natural prosody and realistic lip-smack sounds that define modern conversational AI.

https://deepmind.google/discover/blog/wavenet-a-generative-model-for-raw-audio/