Multi-Stage Neural Audio Processing for Embedded Complexity Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based audio processing methods face challenges due to high computational and memory complexity, making them unsuitable for deployment on embedded devices, and existing input signal normalization techniques, such as power law compression applied only to magnitude, do not effectively address the dynamic range issues of input features.
Innovation Solution
A multi-stage processing approach is introduced, where features are decomposed into subsets with overlapping ranges, processed by neural networks of varying complexities, and combined using a low-complexity network, with power law compression applied to all feature components to improve learning and generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning-based audio processing methods are used, then processing performance is improved, but computational complexity and memory requirements increase
Solution Approach 1:
The patent segments the feature processing into multiple stages: first extracting features from audio signals, then processing different feature subsets through separate neural network paths, and finally combining the results. This segmentation allows the system to achieve deep learning performance while reducing the computational burden on any single processing unit, making it suitable for embedded devices.
Solution Approach 2:
The patent dynamically adjusts the processing complexity by selectively applying neural network processing to different feature subsets based on their importance. Less critical features undergo simpler processing while critical features receive more intensive processing, creating a dynamic complexity adaptation that balances performance and computational requirements.
2Ease of manufacture
If power law compression is applied only to magnitude components, then implementation is simple, but dynamic range issues of input features are not effectively addressed
Solution Approach 1:
The patent applies power law compression universally to all feature components (magnitude, real, and imaginary parts) rather than selectively to only magnitude components. This universal application ensures consistent dynamic range handling across all feature types, improving the overall effectiveness of the normalization while maintaining implementation simplicity through a unified approach.
Data Source
AI summary
An apparatus for processing an information signal has: a feature extractor for extracting a set of features from the information signal, each feature having at least two feature components, and the set of features having a first subset with the first feature components and a second subset with the second feature components; and a neural network processor having: a first neural network for receiving, as input, the first subset and outputting a processed first subset; a combiner for combining the processed first subset and the second subset to obtain a combined subset; and a second neural network for receiving, as input, the combined subset and outputting a processed combined output, which represents a processed information signal, or the apparatus being configured to calculate the processed information signal using the processed combined output, and a complexity of the first neural network being greater than that of the second neural network.