Audio Source Separation Using Iterative Wiener Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for source separation in multi-channel audio signals face challenges in achieving real-time processing with low system and analysis delays, particularly in blind source separation, where prior information about the sources is unavailable, and require efficient algorithms to handle practical computation resources.

Innovation Solution

The method employs a flow chart process involving initialization of source parameters, iterative updates of Wiener filter parameters, and non-negative matrix factorization to separate audio sources from multi-channel signals, using covariance matrices and spectral power matrices to reconstruct audio sources in real-time, applicable to both blind and informed source separation scenarios.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If complex source separation algorithms are used to improve separation quality, then manufacturing precision is improved, but device complexity increases and real-time processing becomes difficult

Engineering Contradiction:
Improvesource separation qualityVSAvoidalgorithm complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the source separation problem into multiple frequency bins, processing each frequency component separately through independent Wiener filters. This segmentation allows complex multichannel separation to be broken down into simpler parallel sub-problems that can be solved efficiently in real-time while maintaining high separation quality through frequency-specific optimization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the source separation problem by changing parameters from time-domain to frequency-domain representation using Fourier transforms. This parameter transformation enables the use of frequency-dependent Wiener filters that adapt to the spectral characteristics of different sources, improving separation quality while the structured frequency-domain approach keeps computational complexity manageable through efficient FFT-based implementation.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If look-ahead data is used to improve separation accuracy, then manufacturing precision is improved, but system delay increases

Engineering Contradiction:
Improveseparation accuracyVSAvoidsystem delay
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent implements periodic processing using overlapping analysis windows with fixed hop sizes, creating a structured sequence of processing stages. This periodic approach allows the system to process audio in manageable frames with controlled look-ahead requirements, maintaining separation accuracy through consistent spectral analysis while limiting system delay by processing periodic segments rather than requiring extensive future data.

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The patent performs preliminary spectral analysis and Wiener filter computation for each frequency bin before final signal reconstruction. By pre-computing frequency-domain representations and filter parameters for each frame, the system achieves high separation accuracy through thorough spectral processing while managing overall delay through efficient in-place updates and overlapping window strategies that minimize look-ahead requirements.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If real-time processing is implemented to reduce system delay, then productivity is improved, but analysis delay increases due to computational complexity

Engineering Contradiction:
Improvereal-time processing capabilityVSAvoidanalysis delay
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent replaces complex time-domain convolution operations with efficient frequency-domain multiplication using Fast Fourier Transforms. This substitution of mechanical computation with frequency-domain algebraic operations dramatically reduces computational complexity from O(N²) to O(N log N), enabling real-time processing with minimal analysis delay while maintaining high separation quality through frequency-specific Wiener filtering.

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

Solution Approach 2:

The patent implements continuous real-time processing through overlapping analysis windows and iterative Wiener filter updates that process audio streams without interruption. This continuous action maintains separation quality by constantly adapting filters to changing spectral conditions while achieving low analysis delay through efficient in-place frequency-domain computations and incremental updates rather than requiring complete reprocessing of entire audio segments.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3440670B1Audio source separation
Publication Date: 2022.01.12 DOLBY LABORATORIES LICENSING CORP
  • EP3440670B1 patent drawingFigure 1
  • EP3440670B1 patent drawingFigure 2
  • EP3440670B1 patent drawingFigure 3

AI summary

The present document describes a method (100) for extracting audio sources (301) from audio channels (302). The method (100) includes updating (102) a Wiener filter matrix based on a mixing matrix from a source matrix and based on a power matrix of the audio sources (301). Furthermore, the method (100) includes updating (103) a cross-covariance matrix of the audio channels (302) and of the audio sources (301) and an auto-covariance matrix of the audio sources (301), based on the updated Wiener filter matrix and based on an auto-covariance matrix of the audio channels (302). In addition, the method (100) includes updating (104) the mixing matrix and the power matrix based on the updated cross-covariance matrix of the audio channels (302) and of the audio sources (301), and/or based on the updated auto-covariance matrix of the audio sources (301).