Meta-Learning Lip Generator for Few-Shot Audio Synchronization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for generating 2D animated lip images struggle with synchronization to audio signals, particularly for unseen subjects, requiring large training datasets, lengthy training times, and often produce blurry or temporally meaningless images, with existing models lacking generalization and requiring retraining for each subject.

Innovation Solution

A meta-learning based model-agnostic technique is employed to generate 2D animated lip images, using a system with a meta-trained audio encoder, image encoder, and decoder, which enables quick learning from limited datasets and fine-tuning within a short time frame, utilizing deep speech features for better encoding and synchronization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If recurrent neural network predicts mouth shape for every frame of audio, then lip synchronization is achieved, but the model requires re-training with hours or minutes of video for each new subject, resulting in loss of time and high device complexity

Engineering Contradiction:
Improvelip synchronization accuracyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system performs preliminary action by pre-training a general talking head model on diverse data before deployment. This pre-trained model serves as a foundation that can be quickly adapted to new subjects without requiring extensive re-training, thus resolving the contradiction between maintaining synchronization accuracy and reducing training time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system changes parameters by using a lightweight subject-specific adapter module that learns subject characteristics through few-shot learning. Instead of re-training the entire model, only the adapter parameters are updated with minimal data, achieving both high reliability and fast adaptation.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If personalized network is trained with huge training datasets for each target face, then faithful rendering of lip movement is achieved, but training becomes very time consuming and requires huge computational resources

Engineering Contradiction:
Improvelip movement rendering fidelityVSAvoidtraining efficiency
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The system segments the model into a general talking head backbone and a lightweight subject-specific adapter. The backbone provides general lip synchronization capabilities while the adapter captures subject-specific characteristics. This segmentation allows the majority of the model to remain fixed, requiring minimal training data and computational resources while maintaining high rendering fidelity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system achieves universality by creating a general talking head model that can be adapted to multiple subjects through a single adapter module. This multi-functional approach allows one base model to serve multiple purposes across different subjects, dramatically improving training efficiency while maintaining personalized rendering quality.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Loss of time

If meta-learning based model-agnostic technique is used, then quick learning from limited datasets is achieved, but the system complexity increases due to multiple encoders and decoders

Engineering Contradiction:
Improvefine-tuning timeVSAvoidsystem architecture complexity
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The system introduces an intermediary adapter module that mediates between the general talking head model and subject-specific requirements. This adapter acts as a bridge, enabling quick adaptation to new subjects without requiring complex re-training of the entire system, thus reducing fine-tuning time while managing system complexity through modular design.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3965102B1Method and system for generating 2d animated lip images synchronizing to an audio signal
Publication Date: 2024.07.17 TATA CONSULTANCY SERVICES LTD
  • EP3965102B1 patent drawingFigure 1
  • EP3965102B1 patent drawingFigure 2
  • EP3965102B1 patent drawingFigure 3

AI summary

This disclosure relates generally to a method and system for generating 2D animated lip images synchronizing to an audio signal for an unseen subject. Recent advances in Convolutional Neural Network (CNN) based approaches generate convincing talking heads. Personalization of such talking heads requires training the model with large number of samples of the target person which is time consuming. The lip generator system receives an audio signal and a target lip image of an unseen target subject as inputs from a user and processes these inputs to extract a plurality of high dimensional audio image features. The lip generator system is meta-trained with training dataset which consists of large variety of subjects' ethnicity and vocabulary. The meta-trained model generates realistic animation for previously unseen face and unseen audio when finetuned with only a few-shot samples for a predefined interval of time. Additionally, the method protects intrinsic features of the unseen target subject.