Highly Persuasive Digital Human: Powered by Ernie Multimodal LLM with Script Generation, Ultra-Realistic Voice and Appearance 170

What is the High-Persuasion Digital Human?

The High-Persuasion Digital Human is a digital human product released by Baidu founder Robin Li at the Baidu Create conference on April 25, 2025. The High-Persuasion Digital Human features ultra-realistic voice and appearance, more professional content, and more flexible interaction, delivering an experience that surpasses that of a real person. It is built on Baidu's e-commerce "Huibox Star" platform, powered by the ERNIE 4.5 large model, and equipped with "script generation" capabilities that allow expressions, tone, and movements to adjust in real time with the script, making emotional transitions and action switches natural and smooth. For example, the digital humans generated by Baidu are extremely realistic in appearance, voice, movements, and expressions—almost identical to real people.

Core Technologies of the High-Persuasion Digital Human

• Script-Driven Multimodal Matching: Leveraging large language models to generate scripts, covering the generation of lines, visual tags, and speech tags. When generating lines, information such as the digital human's movements, expressions, and intonation is generated simultaneously, aligning positions across modalities to achieve multimodal content matching and positional alignment, producing expressive digital humans that capture both form and spirit.

• Multi-Agent Scheduling: Based on real-time popularity and conversion data in the livestream, the digital human can flexibly dispatch roles such as co-host, floor manager, and operations staff to drive conversions—for example, switching camera views at the right moment, and pulling up images and video assets—achieving "one person is an entire marketing team." For instance, in an auto livestream, while the digital human host explains, the co-host simultaneously pulls up a 3D car model, boosting conversion efficiency by 40%.

• Deep Multimodal Fusion: It can perfectly replicate both real people and livestream set designs. Baidu has developed a digital human video generation large model that uses multimodal signals such as video, scripts, and speech for highly controllable digital human generation, then generates video content of sufficient length through spatial alignment.

• Speech Synthesis for Digital Human Scenarios: Based on the ERNIE large model, it predicts text style, emotion, and prosodic起伏, and uses a text-controllable speech synthesis large model to achieve instruction-following synthesis, producing smooth and natural voice effects. Combined with livestream text and speaker information, it synthesizes natural and fluent voice output.

Application Scenarios of the High-Persuasion Digital Human

• E-commerce Livestreaming: This is an important application area for the High-Persuasion Digital Human. It can livestream tirelessly, covering the entire process from product explanation to atmosphere guidance. For example, in a beauty livestream, the digital human导购 achieved a 25% higher conversion rate than a real person; "Rongrong's Exquisite Life" used the Huibox Star digital human for livestreaming, reaching a monthly GMV of over 150,000 yuan, with revenue increasing by 30%; "Beautiful Homewear Factory" saw a 60% GMV growth; the beauty and personal care brand "Zuoyan Yose" used digital humans for makeup try-on demonstrations, achieving over 290,000 yuan in single-session GMV and a 60% increase in livestream conversion rate. The digital human can also make automatic decisions based on real-time feedback in the livestream, such as distributing coupons and switching products.

• Other Fields: The digital human system is currently expanding into multiple industries and is widely applied in vertical scenarios such as law, education, finance, and automotive. Some auto dealers use digital humans for vehicle introductions and interactive Q&A; in book livestreams, digital humans can cite classics; when selling health products, they can provide personalized recommendations based on user data such as blood glucose levels. In the legal sector, "Law in the Heart" has added a lawyer digital human project. Relying on Huibox Star's script creation and big-model knowledge reserves, it accurately answers legal questions and avoids risky terms. After launch, the cost per lead decreased by 75% and the lead volume increased by nearly 5 times.

Advantages of the High-Persuasion Digital Human

• Low Barrier to Entry: Baidu's Huibox Star offers a "one-click cloning" feature—users only need to record and upload a live-streaming video as short as 2 minutes for training, and they can have their own exclusive digital human, allowing anyone to become a host. It also uses AI to help write livestream scripts, creating personalized livestream rooms.

• Low Cost: Currently, the average monthly cost of using the "High-Persuasion Digital Human" for livestreaming is controlled within a thousand yuan, saving about 80% compared to traditional livestreaming. System technology costs are also declining, dropping by about two-thirds over the past year. On the Baidu platform, merchants can start livestreaming at almost zero cost with platform support.

Tutorial on Using the High-Persuasion Digital Human

1. Record a Video: Open the Baidu app, search for "Huibox Star", and click "Mobile One-Click Livestream" and "Create Digital Human". Use your phone to record a 2-3 minute video of a real person on camera. The video requires the face to be visible throughout—no professional equipment is needed.

2. Choose or Customize the Persona, Theme, and Script: While waiting for the video to upload, you can choose a persona prepared by Huibox Star, such as emotional counseling or book sharing. The AI will automatically generate a persona guide, livestream theme, and script. You can also customize a persona, such as a fitness enthusiast, and select "AI Helps Me Write" to generate the script. The entire script generation process takes less than 2 minutes.

3. Wait for the Digital Human to Be Created: The production time is generally around 3 hours. Once completed, you will be notified via SMS, and the page will also indicate the livestream start time.