Dynamic Range Compression of Signal Features 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 power law compression techniques only address the magnitude component, not improving learning and generalization capabilities effectively.
Innovation Solution
Applying power law compression to the complete input feature representation, including both magnitude and phase components, and decomposing input features into multiple components with varying importance for multi-stage processing using neural networks with declining complexity.
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 audio signal processing into multiple stages: feature extraction, dynamic range compression, neural network processing, and output generation. This segmentation allows each stage to be optimized independently, reducing overall computational complexity while maintaining processing performance.
Solution Approach 2:
The patent applies dynamic range compression to the input features before they are processed by the neural network. This preliminary action reduces the dynamic range of the input data, which simplifies subsequent neural network processing and reduces computational requirements while preserving the essential information needed for high-quality audio processing.
2Reliability
If deep learning-based audio processing methods are used, then processing performance is improved, but memory requirements increase
Solution Approach 1:
The patent divides the processing pipeline into distinct modules that can be executed sequentially with minimal memory retention between stages. The feature extraction module processes the input, passes compressed features to the neural network, which then generates output without requiring the entire signal to be held in memory simultaneously.
Solution Approach 2:
By applying dynamic range compression before neural network processing, the patent reduces the variance and range of input values. This preliminary transformation allows the neural network to operate with smaller precision requirements and reduces the memory needed to store intermediate activations and gradients during training and inference.
3Ease of manufacture
If power law compression is applied only to magnitude component, then implementation is simple, but learning and generalization capabilities are not improved effectively
Solution Approach 1:
The patent applies different processing treatments to different components of the audio signal. Specifically, it applies dynamic range compression to both the magnitude and phase components of the Fourier transform, recognizing that each component has different characteristics and requires tailored processing to optimize neural network learning and generalization performance.
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, wherein the feature extractor has: a raw feature calculator for calculating raw feature results, each raw feature result having at least two raw feature components; and a raw feature compressor for performing a compression of a dynamic range to the at least two raw feature components to obtain at least two compressed raw feature components for each raw feature result, wherein the set of features has the compressed raw feature components; and a signal processor for processing the set of features to obtain the processed information signal, wherein the information signal has an audio signal, an image signal, or a radar signal.