Unified Sound Source Separation Model for Low-Memory Audio Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing sound source separation technologies using neural networks require extensive learning processing and memory resources for each sound source type, leading to high learning costs and storage demands, making it difficult to implement in devices with limited capacity.

Innovation Solution

A sound source separation processing method that uses a single learning model to extract multiple types of musical instrument sounds, including models for vocal, guitar, piano, and noise sounds, reducing the need for individual neural networks and memory requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a neural network unique to each sound source type is generated, then sound source separation performance is improved, but learning cost and memory capacity requirements increase enormously

Engineering Contradiction:
Improvesound source separation performanceVSAvoidlearning cost and memory capacity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies universality by creating a single neural network model that can separate multiple sound source types (vocal, guitar, piano, noise) instead of requiring separate specialized networks for each sound source. This multi-functional approach reduces the total number of models needed while maintaining separation effectiveness across different sound categories.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent merges multiple sound source separation functions into one unified neural network model. By combining the separation capabilities for vocal, guitar, piano, and noise sources into a single model structure, the system reduces redundancy and lowers both learning costs and memory requirements while preserving the separation performance that would otherwise require multiple specialized networks.

Inventive Principle:
Principle #5Merging (Combining)

2Adaptability or versatility

If the number of sound source types to be separated increases, then separation capability is improved, but learning processing cost becomes enormous

Engineering Contradiction:
Improveseparation capabilityVSAvoidlearning processing time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent creates a universal neural network model that can handle multiple sound source types including vocal, guitar, piano, and noise sources. This single multi-functional model replaces the need for separate learning processes for each sound source type, significantly reducing the total learning processing time while maintaining the ability to separate all specified sound categories.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If unique parameters for each sound source are set, then separation precision is improved, but memory capacity requirements increase

Engineering Contradiction:
Improveseparation precisionVSAvoidmemory capacity
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent combines the parameter sets for multiple sound source types into a single unified parameter structure within one neural network model. Instead of storing separate unique parameters for vocal, guitar, piano, and noise sources, the system uses a consolidated parameter set that achieves precise separation across all sound types, thereby reducing memory capacity requirements.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4708288A1Information processing device, sound source separation processing method, and program
Publication Date: 2026.03.11 SONY GROUP CORP
  • EP4708288A1 patent drawingFigure 1
  • EP4708288A1 patent drawingFigure 2
  • EP4708288A1 patent drawingFigure 3

AI summary

Provided are an apparatus and a method for executing sound source separation processing for extracting a plurality of different musical instrument sounds and the like using one learning model. The apparatus includes a sound source separation processing execution unit that executes sound source separation processing, and the sound source separation processing execution unit includes a configuration that executes sound source separation processing using a learning model, and executes processing of extracting a plurality of different types of musical instrument sounds using a same learning model. The sound source separation processing execution unit includes three types of learning models including (a) a learning model that extracts a vocal sound (singing sound) and a voice sound (conversation sound), (b) a learning model that extracts a guitar sound and a piano sound, and (c) a learning model that extracts a noise sound.