Conditioned Audio Enhancement Neural Network for Multi-Loss Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep learning models for audio enhancement are computationally expensive and inefficient when covering multiple signal categories and conditions, requiring separate neural networks for each category, bitrate, and codec, leading to high computational load and memory usage.

Innovation Solution

A loss conditional training method for a neural network, where a coefficient vector is randomly sampled from a distribution to condition the network, allowing a single neural network to cover a wide range of coefficients and conditions, using Feature-wise Linear Modulation (FiLM) and multi-resolution STFT loss functions to balance loss terms, enabling efficient training and inference across different signal categories and conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If traditional audio processing methods are used, then computational resources are consumed, but processing speed and real-time performance deteriorate

Engineering Contradiction:
Improveprocessing speedVSAvoidcomputational resource consumption
Core Design Contradiction:
SpeedVSUse of energy by moving object

Solution Approach 1:

The patent transforms the audio processing approach by changing the fundamental parameters from traditional spectral analysis to neural network-based feature extraction. This parameter change enables real-time processing by optimizing the computational characteristics of the system, achieving both speed improvement and controlled resource consumption.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces traditional mechanical/audio signal processing methods with a neural network system. This substitution introduces a learning-based approach that can process audio data more efficiently, achieving real-time performance while managing computational resources through optimized network architecture and training techniques.

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

2Measurement precision

If complex audio processing models are used, then processing accuracy is improved, but model complexity and training difficulty increase

Engineering Contradiction:
Improveprocessing accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the audio processing task into distinct components handled by different neural network modules. This segmentation allows each module to specialize in specific aspects of audio processing, improving overall accuracy while managing complexity through modular architecture that can be trained and deployed independently.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic elements in the neural network architecture, allowing the model to adapt its processing characteristics based on the input audio characteristics. This dynamic approach enables the system to maintain high accuracy across diverse audio inputs while managing complexity through conditional processing paths rather than uniformly complex architecture.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4537333B1Use of a loss term conditioned neural network for processing of audio
Publication Date: 2026.04.29 DOLBY INTERNATIONAL AB
  • EP4537333B1 patent drawingFigure 1
  • EP4537333B1 patent drawingFigure 2
  • EP4537333B1 patent drawingFigure 3

AI summary

A computer-implemented method of loss conditional training of a neural network for outputting an enhanced audio signal, the method including: randomly sampling a coefficient vector from a distribution of coefficients, wherein elements of the coefficient vector are indicative of weight coefficients corresponding to loss terms of a loss function; conditioning the neural network based on the coefficient vector; and training the conditioned neural network based on an audio training signal, wherein the training involves calculating the loss function for the audio training signal after processing by the conditioned neural network, using the weight coefficients indicated by the coefficient vector.