Synthesis of group of CST helped to read children's fairy tales

2020: Scoring of fairy tales

Speech synthesis of CST group (enters Sber's ecosystem) is used for scoring of children's fairy tales on MySkazka service now. It became known on November 26, 2020. By means of technology 10 fairy tales which can be reproduced a female and male voice were read.

Speech synthesis is a language translation technology of the printing text in the sounding speech. In group of CST this technology is created on a stack of methods of deep training that allows to achieve the high-quality synthesized voice.

Feature of synthesis of CST — in use of difficult neural network models for continuous generation of a speech audiosignal in the text, deep syntax and lexical analysis of the text, modeling of intonations, a possibility of modeling of breath. It allows to achieve smoothness and expressiveness of the artificial speech, to make the speech of more realistic. Synthesis of group of CST works as a part of difficult products and AI solutions in the different industries through the whole country: in banks, a telecom, medicine, etc. The joint project on integration of synthesis into the MySkazka project — special for us as it is connected with the most young audience and we are glad to support him. It is sure that the project will develop.

The project was started at the end of August, 2020, then there was a question of implementation of a postscoring of fairy tales.

We were faced by a difficult task as in work of service personal variables which the user fills in real time are used. Therefore simple option — to read our fairy tales using the professional announcer, did not suit us. We began to look for the technological solution and selected synthesis of group of CST: for us it was important not just to synthesize the speech from texts of fairy tales, but to make it the most similar to this — with intonation, aspiration, punctuation marks, target audience of MySkazka service — children. Scoring of fairy tales using synthesis will allow to use service to children who are not able to read yet or experience specific difficulties with reading and also that who prefers a format of audiobooks. The postscoring works only a week, but we already observe positive dynamics — Retention rate of service (retention ratio of users) grew by 30%, and conversion of new users in registration increased from 7 to 11%.