Recently, a video of Zhu Guangquan and the "sign language anchor" online pk of the satin player went viral, and the number of plays just online exceeded one million. In the video, Zhu Guangquan challenges the "hand speed" of the "sign language anchor" with a super fast speech speed, and the sign language anchor also easily copes with his excellent professional ability. In fact, the sign language anchor is not a real person, but a virtual digital person.
This digital person is the first AI sign language anchor created by the "Baidu Intelligent Yunxiling" digital person platform. In the ice and snow event that has just been opened, the Baidu Intelligent Cloud AI sign language anchor has been launched, providing 24-hour sign language services for hearing-impaired users, so that they can also quickly obtain event information.
(AI Sign Language Anchor)
According to statistics, about 430 million people worldwide suffer from moderate and above hearing impairment, according to the data of the second national sample survey of disabled people, there are 27.8 million hearing impaired people in China. Although they can't hear the sound of the arena, they also want to feel the excitement of the game. AI sign language digital people have undoubtedly become an important bridge for them to watch the game.
Wu Tian, vice president of Baidu Group, said: "Baidu has released an intelligent digital human platform that integrates digital human production, content creation and business configuration services - Baidu Intelligent YunxiLing, Baidu hopes to shorten the production cycle of virtual people to the hour level, reduce production costs, and use AI-driven content generation through leading full-stack AI capabilities." ”
How hard is it to develop an AI sign language anchor?
The essence of the sign language digital person is to solve the problem of translation.
Different from traditional translation, the difficulty of sign language translation is that sign language is not translated word by word according to the language, but needs to be refined and adjusted according to the overall meaning of the sentence, and it is also necessary to cooperate with expressions and lips in real time to help the hearing impaired better understand. This has led to AI sign language digital people facing extremely complex technical problems in terms of information condensation, low latency and refinement. Therefore, AI sign language anchors not only need to have a high-precision digital human image, but also need to have an AI brain capable of speech recognition, sign language translation and sign language expression.
(AI Sign Language Anchor Face Detail Map)
From the appearance point of view, compared with the previous AI anchor image, the AI sign language anchor created by Baidu Intelligent Cloud not only highly restores the skin of real people, but also greatly improves in three-dimensionality, flexibility and plasticity. It is reported that Baidu Intelligent Cloud has established a large-scale sign language action library, based on the "National General Sign Language Dictionary" specification, combined with the two-way tuning of motion capture equipment and real sign language teachers, refining nearly 10,000 sign language movements, ensuring the accuracy of sign language expression. At the same time, the first 4D scanning technology is created, which makes the expression of AI sign language anchors natural and vivid, and the accuracy of mouth shape generation is as high as 98.5%.
And her "AI brain" relies on Baidu Intelligent Cloud. Through Baidu's self-developed machine translation technology, Baidu Intelligent Cloud has built a set of accurate sign language translation engines, which can be understood by more than 85%, comparable to the mainstream Chinese-English, Sino-Japanese and other directions of machine translation results, reaching the industry-leading level, combined with Baidu's self-developed speech recognition technology, the text and audio and video content of ice and snow events can be quickly and accurately converted into sign language; at the same time, through the natural action engine optimized for sign language, complete the action drive of AI sign language anchors, and interpret the actions of digital people in real time. Expressions and lip talk. This ensures that AI sign language anchors have highly understandable sign language expression capabilities and accurate and coherent presentation effects.
It is understood that such a complex digital human research and development work, Baidu only took less than 2 months. Baidu said that virtual digital people need a strong knowledge system and dialogue ability to provide services for humans and communicate naturally with people. Baidu's knowledge graph has accumulated 550 billion yuan of knowledge, helping digital people to have "their own" experience and understanding to think. Based on PLATO, a 10-billion-parameter dialogue model released by Baidu last year, virtual people can communicate with users with material, interesting and emotional.
Baidu Intelligent Yunxiling Digital Human Platform: Generate 2D digital people in half an hour, and complete 3D digital people in one or two weeks
At present, digital people have not yet achieved large-scale industrial landing, mainly because of the high threshold of digital production technology, long cycle and high cost. To create a high-precision, high-fidelity non-specific virtual person, it often requires millions or even tens of millions of capital investment.
In December 2021, Baidu released an intelligent digital person platform that integrates digital person production, content creation, and business configuration services, "Baidu Intelligent Cloud Xiling", based on four major technology engines of portrait drive, intelligent dialogue, voice interaction, and intelligent recommendation, to achieve digital people's "ability to hear, speak, understand, and interact". At the same time, Baidu Intelligent YunxiLing provides a convenient and simple operation platform to help enterprises quickly customize 2D cartoons, real images, 3D cartoons, and hyper-realistic digital portraits at low cost, realize the whole process of dialogue services and content creation, and greatly reduce the difficulty of digital human development.
Not long ago, Wu Tian, vice president of Baidu Group, mentioned in a guest appearance on CCTV's "Dialogue" program, "Baidu hopes to shorten the production cycle of virtual people to the hour level through leading full-stack AI capabilities, reduce production costs, and use AI to drive content generation." ”
(Baidu Intelligent Cloud Xiling Platform Architecture Diagram)
For example, as long as a user uploads a photo on the Baidu Intelligent Yunxiling platform, it can quickly generate a 2D digital portrait that can be driven by AI, and the 2D digital person who originally took a week to make the production process can now be shortened to half an hour. The 3D digital people that used to take two or three months to make can now be compressed to one or two weeks, and the development of digital people's personalized functions is more perfect than before, such as through 20 sentences of lightweight model training, the voice of digital people can be generated, that is, users can generate digital people of their own timbre when they say 20 sentences.
(Digital person operation background diagram)
At present, Baidu Intelligent Cloud Xiling has provided one-stop virtual hosts, virtual employees, virtual idols, brand spokespersons and other industries for radio and television, mutual entertainment, finance, government affairs, operators, retail and other industries. The cases that have landed include: SPDB wealth management commissioner digital person "Xiaopu", China Unicom digital customer service "Xiao U", Baidu APP spokesperson "Gong Jun digital person", Baidu Group digital person "Xijiajia", Mars rover digital person "Zhu Rong" and so on.
Li Shiyan, head of Baidu AI Human-Computer Interaction Lab, said: "Our goal in the past year or two is to let everyone and every enterprise realize the freedom of digital people through the opening of Baidu's intelligent Cloud Xiling platform. ”
(Baidu Intelligent Yunxi Ling Digital Human Family)
Leifeng Network