System synthesis is already able to use voice, which exactly reproduces human.
A year ago a group of scientists from Microsoft speech and dialog research have improved our speech recognition system, reaching a 5.9-percent error rate.
Previously it was considered the average value for the frequency of errors of the people themselves, but later it was shown that the rate of 5.1 percent is more suitable for people who are able to quote what was heard in the conversation.
To reduce the error rate of the speech recognition system, the researchers from Microsoft have included a number of improvements to their acoustic and language models based on neural network. In addition to the General updates for all components of the system, the size of the dictionary of the switch increased from 30 000 to 165 000.
The most important innovation was the so-called “long-term memory based on the session of dialogue.” This means that the new language model enables the system to use the previous conversation as the history when trying to clearly define a specific phrase. In other words, the system requires much less understand what the conversation is, like sports or politics, and finds it suitable under the subject phrase.
“We are working on training computers to not only decode the words and understand their meaning, writes technical officer of Microsoft. The transition from recognition to speech understanding is the next important frontier for speech technology”.
System Microsoft speech recognition is currently used in such services like Cortana and Speech Translator. The company provided a document that describes the features of the new version of the system.
© 2017 – 2019, z-news.link. All rights reserved.