Neural Network Parameterization for Physical Sound Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for parameterizing physical models in sound generation for music instruments are largely heuristic and dependent on the sound designer's taste, leading to inconsistent sound quality and long realization periods, and are not suitable for objective optimization or extension to multiple physical models.
Innovation Solution
A system that extracts features from raw audio signals using neural networks to estimate parameters for physical models, allowing for objective acoustic metric evaluation and iterative optimization, enabling the selection of the best physical model parameters for synthesized sound generation independent of the model's structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If heuristic parameterization methods are used by Sound Designer, then sound quality can be customized according to music taste, but the realization period becomes long and consistency is poor
Solution Approach 1:
The patent replaces the manual heuristic parameterization process with an automated neural network system. The neural network learns optimal parameter mappings from training data and automatically generates parameters for new sounds, eliminating the need for manual Sound Designer intervention and achieving consistent results across different sounds and designers.
Solution Approach 2:
The system enables self-service by allowing the neural network to automatically parameterize physical models without human intervention. The network takes audio signals as input and directly outputs the corresponding parameters, making the process autonomous and eliminating dependency on Sound Designer expertise and time.
2Extent of automation
If manual parameterization by Sound Designer is used, then sound character reflects designer's expertise, but the process is not objective and lacks reproducibility
Solution Approach 1:
The patent substitutes the subjective human judgment process with an objective neural network-based system. The network uses learned acoustic features and metrics to objectively determine parameters, replacing the non-reproducible human expertise with a consistent, measurable, and repeatable automated process.
3Adaptability or versatility
If single physical model is used, then system complexity is reduced, but adaptability to different sound types is limited
Solution Approach 1:
The patent creates a universal parameterization system that can handle multiple physical models through a single neural network architecture. The network learns model-agnostic representations from training data and can generalize to different physical models without requiring separate specialized systems for each model type.
Solution Approach 2:
The system segments the parameterization task into independent learnable components within the neural network. By training on diverse physical models during the learning phase, the network learns to decompose and reconstruct parameters for different model types, enabling flexible adaptation without increasing overall system structural complexity.
Data Source
AI summary
A generation system (100) of synthesized sound comprises: a first stage (1), wherein features (F) are extracted from an input raw sound and parameters of said features are evaluated; a second stage (2) wherein the evaluated parameters are used to create a plurality of physical models that are metrically evaluated in order to find the parameters of the best physical model, and a third stage (3) wherein the parameters of the best physical model are perturbed in order to create perturbed physical models and a metric evaluated of the perturbed physical models is performed to find the parameters of the best physical model.