Audio-Driven 3D Facial Animation Using Latent Variable Decoupling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional facial animation techniques require motion capture, focus primarily on mouth and lip movements, and use supervised learning approaches that lead to unnatural animations and averaging effects, failing to generalize across diverse speakers in terms of age, gender, and ethnicity.

Innovation Solution

A computer-implemented technique that decouples content and style representations in facial animation, using machine learning to generate latent variables for content and style, including age, gender, and emotion, to synchronize full facial features without motion capture, employing semi-supervised learning and segmenting facial features for more realistic and generalized animations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If motion capture is used to capture facial movements, then the accuracy of facial animation is improved, but the device complexity and cost increase

Engineering Contradiction:
Improvefacial animation accuracyVSAvoidmotion capture device complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent replaces the mechanical motion capture system with an audio-based neural network system. Instead of using cameras and motion sensors to track facial movements, the system uses audio input processed through neural networks (including VAE and GAN models) to generate facial animation, eliminating the need for complex mechanical capture devices while maintaining animation quality

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces audio signals as an intermediary between the speaker and the facial animation system. The audio contains both content information (what is said) and style information (how it is said, including emotions and facial movement characteristics), which the neural network uses to drive facial animation without requiring direct visual capture

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of operation

If supervised learning is used for audio-based facial animation, then the training process is structured and controllable, but the averaging effect occurs and generalization to diverse speakers deteriorates

Engineering Contradiction:
Improvetraining process controllabilityVSAvoidgeneralization across diverse speakers
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent segments the training process into two distinct phases: an unsupervised pre-training phase using VAE to learn speaker-specific characteristics without labels, and a supervised fine-tuning phase using GAN to refine animation quality. This segmentation allows the system to first capture individual speaker traits (avoiding averaging) and then improve overall animation quality, resolving the contradiction between controllability and generalization

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary unsupervised learning to extract speaker-specific style information before applying supervised training. The VAE model pre-trains on audio data to learn latent representations of individual speaker characteristics, which are then used as input to the GAN for supervised fine-tuning. This preliminary action preserves speaker identity while enabling controlled training

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If conventional techniques focus on mouth and lip movements, then the lip-sync accuracy is improved, but the naturalness of full facial animation deteriorates

Engineering Contradiction:
Improvelip-sync accuracyVSAvoidfacial animation naturalness
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent merges the modeling of mouth/lip movements with other facial features (eyes, eyebrows, head pose) into a unified neural network framework. The audio-driven GAN simultaneously generates coordinates for multiple facial landmarks, ensuring that lip-sync accuracy and overall facial naturalness are achieved together rather than in isolation

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3874460B1Computing system for expressive three-dimensional facial animation
Publication Date: 2024.01.17 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3874460B1 patent drawingFigure 1
  • EP3874460B1 patent drawingFigure 2
  • EP3874460B1 patent drawingFigure 3

AI summary

A computer-implemented technique for animating a visual representation of a face based on spoken words of a speaker is described herein. A computing device receives an audio sequence comprising content features reflective of spoken words uttered by a speaker. The computing device generates latent content variables and latent style variables based upon the audio sequence. The latent content variables are used to synchronized movement of lips on the visual representation to the spoken words uttered by the speaker. The latent style variables are derived from an expected appearance of facial features of the speaker as the speaker utters the spoken words and are used to synchronize movement of full facial features of the visual representation to the spoken words uttered by the speaker. The computing device causes the visual representation of the face to be animated on a display based upon the latent content variables and the latent style variables.