3D digital-human production
Interactive characters are produced around 3D virtual humans: a publisher IP, a figure in the book, and a classroom teacher.
Technology · Stack
Digital humans, large models, and speech carry the interactive layer of a book
The stack is for making new content and for connecting publishing to ongoing digital updates. A joint artificial-intelligence laboratory with the University of Electronic Science and Technology of China is the research side of this work.
Real-time interaction framework
Real-time interaction built on large models. The character follows text and speech and keeps a readable face and mouth while talking.
Cloned digital-human video
Video built on a cloned presenter. Recommendations, introductions, and chapter explanations can be produced in volume with the same face.
BingoMeta Engine
Text drives speech and animation directly
The engine connects a large-model conversation to an interactive digital human. Response stays live. Expression and style extend across a classroom explanation, a story character, or a dialect tour.
TTSA character model
Text to Speech & Animation. One passage of text generates speech and the matching animation, without recording first and then fitting the mouth.
High-fidelity lip sync
Lip motion and animation rendering work with a reconstructed 3D face mesh and texture identity, so the character remains the same person while speaking.
Live facial animation
Speech drives a cartoon or a more realistic character. Facial animation is modeled so emotion changes transition across the face.
Personalization from a small sample
A new publisher IP or teacher can be tuned from a small sample, without a long capture cycle from scratch.
Controllable emotion
A speech-emotion module adjusts style and intensity. The same character can explain calmly or tell a story with more rise and fall.
More than one setting
The same interaction framework is used for recommendations, reading along, classroom aid, and cultural tours. What changes is persona, knowledge, and tone.
BingoEDU Model
Train one book into an agent for its own domain
BingoEDU is a vertical model service for education publishing. The base uses DeepSeek V3 for expert tasks and R1 for reasoning. Above that are the knowledge base, orchestration, and configurable agents.
ApplicationsReview assistant, writing assistant, character dialogue, AI assistant, expert Q&A, and an agent defined for a specific book.
Orchestration and retrievalVisual flow design, retrieval tests, and GraphRAG turn a book’s chapters into a callable flow.
Models and tuningLarge models, embeddings, tuning datasets, and model management. An application can be defined and maintained.
Knowledge and dataDocuments are cleaned, prepared, and batched into vector storage. Multimodal data covers images, text, audio, and video.
An agent for the book
Each title binds its own documents and character settings. Questions are about the people, ideas, and exercises in that book, and answers can be traced to a chapter.
Editors have assistants too
Writing and review assistants face the production side. Updates in fusion publishing are not limited to the reader.
The model and the digital human are one path
The model decides what is said. BingoMeta decides how it is said and how the face moves. The reader sees a person who can explain this book.
Technology is delivered with the book
The digital human, video, live interaction, and vertical model are combined for the readers and the setting of that title.