Jimu News reporter Chen Hong
"Hello everyone, long time no see, I'm your old friend Brother Dong......" In the face of all netizens shouting to start the live broadcast, "Liu Qiangdong" met with everyone in a new form.
At 6:18 p.m. on April 16, the "Procurement and Sales Dongge" AI digital human created by JD Yunyanxi made its live broadcast debut, and at the same time appeared in the live broadcast room of JD Home Appliances and JD Supermarket. The day before, Jingdong officially released a warm-up video, and many netizens were sure that it was not AI in the video, "So many small actions, it must be a real person!"
How is the AI digital human trained in "Procurement and Marketing Dongge"?
"Brother Dong" occasionally rubbed his fingers when he spoke
In the live broadcast debut on April 16th, the AI digital human of "Procurement and Sales Dongge" changed Liu Qiangdong's previous perseverance and hard work style, allowing netizens to see another side of him: talking and laughing freely, talking about his experience in sports and cooking, and also gushing about the healthy combination of large-screen TVs and three meals a day with live broadcasts, becoming the "chief procurement and sales" of Jingdong Live on the same day.
Many netizens were amazed that the AI digital human of "Purchasing and Selling Dongge" has almost 100% restored Dongge's expression, posture, gestures, and timbre, and even a digital clone can have fresh vitality. Not only to "resemble", but also to "resemble", how does the digital human as an imitation and extension of the human image?
"The 'Procurement and Marketing Dongge' AI digital human is made by JD Yunyanxi. The large model will pay attention to capturing and presenting Dong Ge's habitual expressions and movements, such as occasionally rubbing his fingers when speaking, accommodating more large hand movements when emphasizing something, and nodding his head from time to time. The relevant person in charge of Jingdong said,AIThe digital person is actually more challenging is the sound restoration,People who have listened to Dongge's speech,Impressed by his Suqian accent,You will find that he speaks faster,Speak softly,Some words will be used to reading continuously,He is "time""Is""sh"Pronounced with a heavy nasal, and likes to call "brothers" to encourage everyone's morale......
It is understood that it is not difficult to make a digital person speak Mandarin well, but it is not easy to learn my pronunciation habits to make the AI digital person of "Brother Dong" speak "Suqian Mandarin" well. For example, whether to add nasal sounds, where to read continuously, the large model needs to give accurate judgments. Behind this, it relies on the continuous learning and training optimization of the image and voice of the Jingdong Yanxi model.
The relevant person in charge of JD also revealed that in order to make the sound more suitable for live broadcasting, and to use the mantra of "brothers" in the right place, the JD Yunyanxi team has optimized the model for these details. At first, they "fed" the speech material to the large model, which was passionate and explosive, but too formal. To this end, they used the latest recorded chats as the main material, including Dong Ge's vivid travel experience, and then extracted the rhythmic characteristics of the 5-minute speech and poured it into the large model, and through continuous optimization, finally shaped the voice of the "Dong Ge" AI digital human who is friendly, natural, and close to the user.
50,000 hours of voice data are "fed" during training
JD Cloud Yanxi Digital Human has served more than 4,000 brand live broadcast rooms
Behind the birth of the AI digital human of "Procurement and Marketing Dongge", it is inseparable from the deep technical cultivation and accumulation of JD Yunyanxi for many years.
It is understood that during the training, the Yanxi voice model was "fed" with 50,000 hours of massive and fresh voice data, which allowed the Yanxi digital human to intelligently match different live broadcast styles, such as creating a professional atmosphere with a calm timbre, or using a highly infectious voice to attract users to place orders, and also giving Yanxi a physical performance. Experiments have shown that the vast majority of users will not be able to detect that this is a digital human for 120 seconds.
On the basis of zero configuration, Yanxi Digital Human covers 70% of the common inquiries in the live broadcast room, and can also iterate on itself with the help of the large model intelligent Q&A tuning assistant. On the one hand, Q&A is automatically generated according to the business details page and script, and on the other hand, the knowledge points that are not covered are automatically filled in after the live broadcast is over, and a large number of inquiries can be answered efficiently and in real time. For example, like a senior shopping guide who is proficient in business, when someone asks "is there a road bike suitable for girls" or "is there a mobile phone suitable for college students", he can give suitable product recommendations, and the response accuracy rate is over 90%.
Surprisingly, if there are out-of-stock products during the live broadcast, the live broadcast center console will be like the "brain" of the digital person, adjusting the live broadcast skills in time, such as skipping the out-of-stock products, or increasing the frequency of explanations for popular products. It can also monitor whether the interaction in the live broadcast room is too frequent, and adjust the interaction frequency and mechanism to ensure smooth progress.
It is reported that at present, JD Cloud Yanxi Digital Human has served more than 4,000 brand live broadcast rooms. Since July last year, Jingdong released the Yanxi model, based on it, Jingdong has successively launched nearly 100 innovative applications, including Jingdong intelligent shopping guide assistant "Jingyan", medical and health model "Jingyi Qianxun", Jingdong Logistics Superbrain, Jingdong intelligent customer service, Jing Xiaozhi, Yanxi multimodal digital human, through the full-stack technological innovation from the underlying computing power to model services and AI platforms, service platform operation, professional category knowledge enhancement, consumer experience optimization, decision-making cost reduction, intelligent search and push, merchant conversion and other scenarios.
Baidu founder Robin Li appeared on the same day because of AI
Coincidentally, also on April 16, Baidu founder Robin Li was once again active in the public eye, standing for his own technology applications. At the Create2024 Baidu AI Developer Conference, Robin Li emphasized an important trend in the development of AI applications in the future, and released a number of AI development tools, giving the judgment that "open source models will fall behind more and more".
At the meeting, Robin Li delivered a keynote speech on "Everyone is a Developer", he said that AI is setting off a creative revolution, and developing applications in the future is as simple as shooting a short video, everyone is a developer, and everyone is a creator. At the meeting, Baidu officially released the tool version of Wenxin Model 4.0. Users can experience the code interpreter function on the tool version, realize the processing and analysis of complex data and files through natural language interaction, and also generate charts or files, which can quickly gain insight into the characteristics of the data, analyze the change trend, and provide efficient and accurate support for subsequent decision-making.
Robin Li said that large language models themselves do not directly create value, and AI applications developed based on large models can meet real market demand. ”
(Source: Jimu News)
For more exciting information, please download the "Jimu News" client in the application market, please do not reprint without authorization, welcome to provide news clues, and pay the remuneration once adopted. 24-hour hotline: 027-867777777.