Neural Network Sub-Units for Sound Source Separation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep neural network (DNN) based sound source separation techniques require a large number of operations and memory for high performance, leading to increased computational load.

Innovation Solution

A program and information processing method that utilizes a DNN with a multi-layered structure, including an encoder and sub-neural network units, to reduce the dimensionality of feature vectors and optimize operations, thereby minimizing the number of coefficients and operations while maintaining sound source separation performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a deep neural network with a larger number of coefficients is used to achieve high sound source separation performance, then the sound source separation accuracy is improved, but the memory capacity required for storing coefficients increases

Engineering Contradiction:
Improvesound source separation accuracyVSAvoidmemory capacity for coefficients
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent divides the deep neural network into multiple independent sub-neural network units (first sub-neural network unit, second sub-neural network unit, etc.), each processing different feature vectors. This segmentation allows the system to achieve high separation accuracy through coordinated processing of multiple specialized units rather than requiring a single large network with excessive coefficients, thereby reducing overall memory requirements while maintaining performance.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If a deep neural network with a larger number of coefficients is used to achieve high sound source separation performance, then the sound source separation accuracy is improved, but the amount of operations such as multiplication and addition increases

Engineering Contradiction:
Improvesound source separation accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the computational workload across multiple sub-neural network units, each handling specific feature vector processing tasks. This segmentation enables parallel processing of different audio features, reducing the sequential operation count while achieving high accuracy through the collective output of multiple specialized units.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the problem from a single large-network approach to a multi-unit parallel architecture, effectively adding a dimensional aspect to the computation. By processing multiple feature vectors through separate sub-units simultaneously, the system achieves high accuracy without proportionally increasing the total number of operations, as each sub-unit operates on reduced-dimensional input spaces.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Reliability

If the capacity of memory for storing coefficients is increased to use more coefficients, then the sound source separation performance is improved, but the device complexity and resource requirements increase

Engineering Contradiction:
Improvesound source separation performanceVSAvoidsystem resource requirements
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the coefficient storage requirement across multiple sub-neural network units, each with its own optimized set of coefficients. This segmentation allows the system to achieve high performance through distributed, specialized coefficient sets rather than requiring a single large coefficient matrix, thereby reducing peak memory requirements and simplifying the overall system architecture.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4365897B1Computer program, information processing method and information processing device
Publication Date: 2025.06.25 SONY GROUP CORP
  • EP4365897B1 patent drawingFigure 1
  • EP4365897B1 patent drawingFigure 2
  • EP4365897B1 patent drawingFigure 3

AI summary

For example, the number of operations is reduced without a deterioration in sound source separation performance. A program according to the present disclosure causes a computer to execute an information processing method, the information processing method including generating, by a neural network unit, sound source separation information for separating a predetermined sound source signal from a mixed sound signal containing a plurality of sound source signals, transforming, by an encoder included in the neural network unit, a feature extracted from the mixed sound signal, inputting a process result from the encoder to each of a plurality of sub-neural network units included in the neural network unit, and inputting the process result from the encoder and a process result from each of the plurality of sub-neural network units to a decoder included in the neural network unit.