Source Separation via Mixed Multivariate PDF for Frequency Domain ICA

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing frequency domain independent component analysis (ICA) methods for source separation in audio signals face challenges such as permutation problems, increased processing time, and poor performance in noisy environments, especially when dealing with multi-source speech signals and time-varying statistical properties.

Innovation Solution

The use of mixed multivariate probability density functions that account for the relationships between frequency bins, allowing for time-dependent and source-dependent parameter weighting, which helps in accurately aligning frequency bins and improving performance across wider time frames and multiple speakers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If frequency domain ICA is used for source separation, then computational efficiency is improved, but permutation problems and poor performance in noisy environments occur

Engineering Contradiction:

Improvecomputational efficiencyVSAvoidperformance accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent applies dynamics by making the probability density function parameters time-dependent and source-dependent. The mixed multivariate PDF allows parameters to adapt and change over time, capturing the non-stationary statistical properties of speech signals in different time frames and for different sources, thereby improving performance accuracy while maintaining computational efficiency through frequency domain processing

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent changes parameters by using mixed multivariate probability density functions with time-dependent and source-dependent parameters. This allows the statistical model to adapt to varying signal characteristics across different time frames and sources, resolving the permutation problem and improving separation performance in noisy environments while keeping the computational approach efficient

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If conventional ICA is used for simplified instantaneous mixtures, then separation accuracy is improved, but it cannot handle complex real-world mixing processes with noise and reverberations

Engineering Contradiction:

Improveseparation accuracyVSAvoidhandling capability of complex environments

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent extends ICA from static instantaneous mixture assumptions to dynamic convolutive mixture handling by using time-dependent probability density functions. This allows the model to adapt to complex real-world scenarios including noise, reverberations, and moving sources while maintaining separation accuracy through the flexible statistical framework

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent creates a universal source separation framework that handles both simple instantaneous mixtures and complex convolutive mixtures with noise and reverberations. The mixed multivariate PDF approach provides a unified solution that works across different mixing conditions, making the method versatile for various real-world applications

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Reliability

If time domain ICA is used for convolutive mixtures, then source separation is achieved, but computational resources and processing time are excessively high

Engineering Contradiction:

Improvesource separation capabilityVSAvoidcomputational resources

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies dynamics by using time-dependent probability density functions that adapt to changing signal statistics. This allows frequency domain processing to effectively handle convolutive mixtures with the same reliability as time domain methods, but with significantly reduced computational resources by avoiding full time domain deconvolution

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent substitutes the computationally intensive time domain mechanical processing with frequency domain statistical processing. By using mixed multivariate PDFs in the frequency domain, the method achieves source separation capability comparable to time domain ICA but with much lower computational cost, effectively replacing the heavy mechanical time domain processing

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS8886526B2Source separation using independent component analysis with mixed multi-variate probability density function

Publication Date: 2014.11.11 SONY INTERACTIVE ENTERTAINMENT LLC

AI summary

Methods and apparatus for signal processing are disclosed. Source separation can be performed to extract source signals from mixtures of source signals by way of independent component analysis. Source separation described herein involves mixed multivariate probability density functions that are mixtures of component density functions having different parameters corresponding to frequency components of different sources, different time segments, or some combination thereof.