The legendary "reading lip surgery" finally has specialized learning software.

The Verge Chinese Station reported on November 9

“Reading Lips” is a technical activity. Foreign testing experiments have found that most people can only distinguish one tenth of the words by watching the latter's lip movements when others say it, even the so-called lip-speaking experts. Its accurate recognition rate is also not ideal. However, researchers at Oxford University say that artificial intelligence techniques such as deep learning can help solve such problems.

As we all know, the artificial intelligence technology that seeks "common ground" by acquiring a large amount of data can improve the audio language recognition to the same accuracy as the "face to face" dialogue. Why can it not accomplish the task of "reading lip"?

Researchers at Oxford University's artificial intelligence laboratory mentioned in a recent paper that they developed deep-learning software called “lip-reading” software. Their software is called “LipNet” and its “performance” is much better For those lips interpreters: In some tests, LipNet software was able to achieve 93.4% accuracy, while the lips interpreter's accuracy rate was only 52.3%.

Even though it is still in its early stages, the software is running very fast, and it has almost reached a processing speed that can convert silent video into a text script “in real time”.

The researchers used a set of databases to train and test the system. During the test, the researchers collected short videos recorded by 34 volunteers. In the video, the volunteers read some “meaningless” sentences (such as captions), each short video is only three seconds long, and each sentence uses a very simple sentence structure: command verb + color + Preposition + letter + number + adverb, such as "set blue by A four please" or "place red at C zero again".

In fact, these sentences have their limitations. For example, they only used four different instructions and color words. This also led to the questioning of other researchers in the field. They believe that the study report is too watery. Unconvincing.

However, this is not the case. In an interview, the author of the report and two researchers, Yannis Assael and Brendan Shillingford, admitted that their research was limited by the restrictions on words and grammar. However, this is due to the limited data available, the database is very small, but the test results also show that they can perform equally well in larger databases. ”

Both Assael and Shillingford emphasize that their research results are applied to the monitoring field. The reason is very simple. “Reading lips” requires you to look at the mouth of the target person. This means that the camera must be placed in the best position to get good results. result. "From a technical point of view, it is very, very difficult to apply lip gloss in monitoring areas," Assael said.

However, the two researchers said that reading lip artificial intelligence can help those who are hearing impaired, especially in a relatively noisy environment (that is, the computer is difficult to separate the noise environment).

For example, such people can wear glasses with built-in cameras. They can clearly capture the lip movements of the target person when they are at a party, and then use this software to translate the lips “language” into text in real time, and then Speech is transmitted to the wearer's ear.

"As long as you have speech recognition and cameras, we can improve it," Assael said. He also mentioned that Apple Siri or Google Now Voice Assistant will be able to apply their software.

In the future, perhaps we will not dare to speak to our computers. The reason is simple. They may read what we say. (Original author James Vincent compilation: Newsboy)

Wonderful video:

Click to view original english

The Chinese related rights of the works of The Verge in the United States belong to Tencent Corporation and may not be reproduced or excerpted without authorization.

22V AC DC Switching Power Adapter

22V Plug in AC/DC switching power supply were widely used for any small power device, such as CCTV Cameras, wireless routers, LED strip, ADSL cats, HUB, switches, security cameras, audio/video power supply. For 22V wall mount power supply, the maximum output current is 1.64A,total 36W output. Our power adaptor meets different certificates for different countries` request – like UL list/CCC-CQC/ PES/SAA/C-TICK/CB/GB certificate. All our switching power supplies were getting 100% full-load burning test for at least 2 hours, and 3000Vac withstanding voltage test for 1 minutes.

22V1A US Plug In Power Supply

22V AC Switching Power Adapter,22V DC Switching Power Adapter,AC DC 22V 1A Power Supply Switch Universal Power Adapter

Shenzhen Juyuanhai Electronic Co., Ltd. , https://www.powersupplycn.com