Universal Audio Model for Sound Decomposition Without Source-Specific Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional sound decomposition techniques rely on isolated training data from actual sound sources, which can be labor-intensive and resource-consuming, and fail when such data is not available.
Innovation Solution
A universal audio model is generated from a plurality of different sound sources, allowing for sound decomposition without requiring specific training data by selecting models that correspond to the sound data, using techniques like non-negative matrix factorization and block sparsity to guide the decomposition process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional sound decomposition techniques use isolated training data from actual sound sources, then decomposition accuracy is improved, but labor and resource requirements increase significantly
Solution Approach 1:
The patent creates a universal audio model that can decompose multiple types of sound sources (speech, music, noise) using a single unified framework rather than requiring separate training data for each source type. This universal model achieves decomposition accuracy comparable to conventional methods while eliminating the need for extensive isolated training data collection and processing for each sound category.
2Productivity
If conventional techniques are used when training data is not available, then resource requirements are reduced, but decomposition performance deteriorates or becomes impossible
Solution Approach 1:
The patent performs preliminary training on diverse audio data to build a comprehensive universal audio model before actual decomposition tasks. This pre-trained model contains learned representations of various sound sources that can be directly applied to new decomposition problems without requiring additional source-specific training data, ensuring reliable performance even when training data is unavailable.
3Productivity
If a universal audio model is used instead of isolated training data, then resource requirements and labor are reduced, but model complexity increases
Solution Approach 1:
The patent transforms the approach by changing from storing and processing numerous isolated training datasets to a single universal model with learned parameters from diverse data. This parameter transformation consolidates complexity into the model's internal representations rather than external data management, reducing resource requirements while maintaining decomposition capability.
Data Source
AI summary
Sound decomposition models are described. In one or more implementations, a plurality of individual models is generated for respective ones of a plurality of sound sources. The plurality of models is collected to form a universal audio model that is configured to support sound decomposition of sound data through use of one or more of the models. The plurality of models is not generated using a sound source that originated at least a portion of the sound data.