Google search giant specialists published an article in which they talked about creating a speaker generator that can speak and not be different from a living person in a voice. The development is called Tacotron 2 and is able to translate text into speech in a very high quality.

The program is two interrelated neural networks of in-depth learning. The first neural network creates on the basis of the text of the spectrograph and transmits it to the next WaveNet algorithm, which it and voice “voice”. Tacotron 2 knows a lot of nuances, effortlessly cope with complex words in spelling and, when reading from a sheet, takes punctuation into account. Because of this, for example, he distinguishes between the end of the sentence and the beginning of the new one, distinguishing it with intonation.

Examples of the work of the program experts have already laid out on the development page. It sounds much better than the monotone mechanical voices of modern voice recording software, so Google probably will quickly find application development. WaveNet is already used by Google Assistant, so Tacotron 2 will definitely be a great addition to it.

At this stage of development, Tacotron 2 speaks only in a pleasant female voice, but probably in the future will receive a male version, and, given its ability to learn, maybe learn and imitate other voices.