Audio Feature Segmentation and Combination for Embedded Neural Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based audio processing methods face high computational and memory complexity, making them unsuitable for deployment on embedded devices, and existing input value compression techniques, such as power law compression applied only to magnitude, do not effectively address the dynamic range issues in neural networks.
Innovation Solution
A method involving feature segmentation and combination using neural networks, where features are split into subsets with overlapping ranges, processed by neural networks of varying complexities, and combined using a lower complexity network, with power law compression applied to all components of the input feature representation.
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 independent pathways: magnitude feature processing and phase feature processing. Each pathway uses separate neural network models with optimized architectures, allowing parallel computation and reducing the computational burden of a single large model while maintaining overall processing performance.
Solution Approach 2:
The patent extracts and separates the magnitude and phase components of the audio signal for independent processing. By taking out the magnitude features and phase features as distinct processing streams, the system can apply specialized processing to each component, reducing the complexity of processing the complete complex-valued signal as a whole.
2Object-generated harmful factors
If power law compression is applied only to magnitude components, then dynamic range is reduced, but phase components remain uncompressed causing learning issues
Solution Approach 1:
The patent applies power law compression universally to both magnitude and phase components of the audio signal. This multi-functional application ensures that all feature components benefit from dynamic range compression, enabling the neural network to learn effectively from both magnitude and phase information without the limitations of partial compression.
3Reliability
If a single complex neural network processes all features, then comprehensive processing is achieved, but model complexity and training difficulty increase
Solution Approach 1:
The patent divides the neural network processing into segmented pathways: one for magnitude features and one for phase features. Each pathway contains its own neural network model with architecture optimized for the specific characteristics of that feature type, replacing a single complex model with multiple specialized models that are easier to train and compute.
Solution Approach 2:
The patent transforms the processing from a single complex-valued dimension to separate real-valued dimensions for magnitude and phase. This dimensional transformation allows the use of standard real-valued neural network operations on complex audio features, simplifying the computational framework while maintaining processing completeness.
Data Source
AI summary
An apparatus for processing an information signal has: a feature extractor for extracting a set of features having a first dimension; a feature segmenter for segmenting into a first subset having a second dimension and a second subset having a third dimension, which overlap, both being lower than the first dimension; a neural network processor for processing the first and second subsets using a first and a second neural network to obtain a first and a second result, respectively; a feature combiner for combining the first and second results using a third neural network, having a third complexity lower than a first or a second complexity of the first and second neural network to obtain a result set of features having a result dimension; and an output post-processor for post-processing the result set of features to obtain a processed information signal.