Emotion-Editable Digital Human Generation From Speech Encoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current digital human generation technologies lack the ability to generate digital humans with emotions, limiting their effectiveness in interactive scenarios such as product service introductions, medical rehabilitation, and autonomous driving.

Innovation Solution

A method involving encoding input speech and target emotion information to obtain speech and emotion code matrices, fusing these matrices, and inputting the fusion matrix into a digital human generation model to generate a digital human image with emotions, allowing for emotion editing based on scenario needs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If current digital human generation methods (GAN or NeRF) are used, then high quality digital humans with aligned mouth shape and voice can be generated, but digital humans with emotions cannot be generated

Engineering Contradiction:
Improveemotional expression capabilityVSAvoidemotion alignment accuracy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the emotion generation process into distinct components: emotion type identification (categorizing emotions into discrete classes), emotion intensity measurement (quantifying emotional strength), and separate processing streams for different emotional dimensions. This segmentation allows the system to handle complex emotional expressions through modular, manageable stages, enabling emotional capability without compromising generation reliability

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an emotion code matrix as an intermediary representation that bridges speech input and digital human generation. This intermediate structure encodes both emotion type and intensity information in a structured format that can be seamlessly integrated into existing generation pipelines, allowing emotional expressions to be incorporated without disrupting the proven GAN/NeRF generation processes

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of operation

If digital humans can only express in voice, then generation simplicity is maintained, but user interactive needs cannot be met

Engineering Contradiction:
Improvegeneration process simplicityVSAvoidinteractive capability
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamic emotional expression by allowing the digital human's emotional state to change in real-time based on speech content and context. The system dynamically adjusts emotion type and intensity parameters during generation, enabling the digital human to adapt its expressions to match the interactive context while maintaining a relatively simple underlying generation framework

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent performs preliminary emotion analysis on the input speech to identify emotion type and intensity before the actual digital human generation process. This preliminary action prepares emotion codes in advance, allowing the main generation system to remain simple while incorporating pre-computed emotional information that enhances interactive capability

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4668215A1Digital human generation method and apparatus, computer-readable storage medium, and terminal
Publication Date: 2025.12.24 UNIDT (SHANGHAI) CO LTD
  • EP4668215A1 patent drawingFigure 1
  • EP4668215A1 patent drawingFigure 2~3
  • EP4668215A1 patent drawing

AI summary

A digital human generation method and apparatus, a computer-readable storage medium, and a terminal, the method comprising: determining input speech and target emotion information (S11); encoding the input speech to obtain a speech encoding matrix, and encoding the target emotion information to obtain an emotion encoding matrix (S12); fusing the speech encoding matrix and the emotion encoding matrix to obtain a fusion matrix (S13); and inputting the fusion matrix into a digital human generation model to obtain a digital human image corresponding to the input speech (S14). By using the foregoing solution, a digital human with emotion can be generated, and emotion editability of the digital human is achieved.