Blind Source Separation Using Compressibility-Based Clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current single-channel blind source separation methods face challenges in effectively separating source signals from mixture signals, particularly in music, due to limitations in clustering algorithms and compression techniques, leading to suboptimal source separation results.

Innovation Solution

The method employs non-negative matrix factorization (NMF) combined with a new clustering criterion based on compressibility, using linear predictive coding (LPC) error to minimize compression error and Wiener filtering, which allows for improved source separation by grouping spectral components in the time domain, enhancing the accuracy of source signal estimation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional clustering algorithms are used for source separation, then the separation process is simple to implement, but the source separation results are suboptimal

Engineering Contradiction:
Improvesource separation accuracyVSAvoidclustering algorithm complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the clustering criterion from traditional distance-based metrics to compressibility-based metrics (LPC error). By evaluating how well different clusterings compress the separated channels, the method identifies optimal groupings that maximize source separation accuracy. This parameter change transforms the clustering objective from geometric proximity to signal compressibility, resolving the contradiction between simplicity and accuracy.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces traditional mechanical/clustering-based source separation mechanisms with a compression-based evaluation system. Instead of relying on fixed clustering algorithms to group frequency bins, the system uses linear predictive coding error as a feedback mechanism to dynamically determine optimal clusterings. This substitution of the evaluation mechanism enables higher accuracy without proportionally increasing complexity.

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

2Reliability

If compression techniques are applied to improve source separation, then the signal-to-distortion ratio improves, but the computational complexity increases

Engineering Contradiction:
Improvesignal-to-distortion ratioVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies the self-service principle by using the compressibility of the separated channels themselves as the evaluation criterion. The separated signals are compressed using linear predictive coding, and the compression error directly indicates clustering quality. This self-evaluation mechanism eliminates the need for external validation metrics or complex optimization functions, improving reliability while keeping computational complexity manageable.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent implements feedback by using compression error as a performance metric that guides the clustering optimization process. The LPC error provides continuous feedback on how well different clusterings separate the sources, allowing the system to iteratively improve source separation reliability. This feedback mechanism transforms compression from a post-processing step into an active optimization guide.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3201917B1Method, apparatus and system for blind source separation
Publication Date: 2021.11.03 SONY GROUP CORP
  • EP3201917B1 patent drawingFigure 1
  • EP3201917B1 patent drawingFigure 2
  • EP3201917B1 patent drawingFigure 3

AI summary

A method comprising decomposing a magnitude part of a signal spectrum of a mixture signal into spectral components, each spectral component comprising a frequency part and a time activation part; and clustering the spectral components to obtain one or more clusters of spectral components, wherein the clustering of the spectral components is computed in the time domain.