[an error occurred while processing the directive]
RSS
Логотип
Баннер в шапке 1
Баннер в шапке 2

Google Parrotron

Product
Developers: Google
Date of the premiere of the system: July, 2019
Branches: Internet services
Technology: Speech technologies

2019: Announcement

In the middle of July, 2019 Google announced the software which helps people fully to communicate with violations of the speech. The tool received the name Parrotron.

He uses the deep neural network trained to transform the atypical speech to free and clear. The technology converts a voice into the text and back, without resorting to recognition of lip movement and other visual signals.

Google provided software for free communication of people with violations of the speech

Google connected audiomaterials with a total duration of 30 thousand hours which contain millions of anonymous fragments of a talk to a neuronet. According to developers, Parrotron considerably reduces quantity of errors at recognition of the speech for deafs — the probability of an error was reduced from 89% to 25%. Google does not intend to be satisfied with what has already been achieved and promise to finish technology.

According to the VentureBeat edition, by July, 2019 millions of people have violations of the speech worldwide. In the USA such deviations are noticed at 7.5 million people, about 5% of the American first graders have problems with height, volume and quality of sound pronunciation. The variety of deviations complicates to developers of AI systems understanding of a question and creation of the systems of recognition and speech synthesis. They should adapt the technologies for deviations on which small data sets are available to training.

File:Aquote1.png
Parrotron facilitates to users with the atypical speech an opportunity to communicate with other people and to be the understood by them and voice interfaces. For this purpose use a complex method of conversion of the speech which more likely reproduces the expected speech of the user … And as Parrotron is not strongly configured on reproduction of words from the predetermined dictionary set, input data for model may contain absolutely new words, a foreign word, names and even senseless words — the researcher Fadi Biadsy and the software engineer RON Weiss working in Google reported.[1]
File:Aquote2.png

Notes