Neural Network Bandwidth Extension for Audio Signal Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing bandwidth extension (BWE) methods in audio encoding require additional bits for encoding side information or have high network complexity, leading to performance issues in scenarios beyond training modes and forward compatibility problems.
Innovation Solution
A BWE method using a neural network model to determine correlation parameters between low-frequency and high-frequency spectra, allowing for bandwidth extension without additional bits and reducing network depth and complexity, by inputting low-frequency spectrum parameters into the model to obtain a correlation parameter and generate a high-frequency spectrum.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a classic BWE method using correlation as side information is used, then bandwidth extension can be achieved, but additional bits are required for encoding and forward compatibility is compromised
Solution Approach 1:
The patent extracts only the essential correlation information needed for BWE and integrates it directly into the neural network model's internal representations, removing the need to transmit separate side information bits. The network learns to embed correlation patterns within its hidden layer activations rather than requiring explicit bitstream transmission.
Solution Approach 2:
The neural network model serves multiple functions simultaneously: it performs both the primary audio signal processing and implicitly captures correlation information needed for bandwidth extension. This multi-functionality eliminates the need for dedicated side information channels, as the same network structure handles both tasks.
2Quantity of substance
If a blind solution based on neural network or deep learning is used, then no additional bits are needed, but the network has large depth and volume with high complexity
Solution Approach 1:
Instead of using a complete blind BWE neural network with full depth and complexity, the patent employs a simplified version that processes only the necessary features for the specific application scenario. The network architecture is trimmed to include only essential layers and neurons required for effective bandwidth extension in practice.
Solution Approach 2:
The patent divides the bandwidth extension task into separate processing stages, each handled by a specialized neural network module. This segmentation allows each module to be smaller and less complex than a monolithic BWE network, while collectively achieving the full bandwidth extension function through modular cooperation.
3Adaptability or versatility
If a neural network with large depth and volume is used, then generalization capability is improved, but performance is mediocre in scenarios beyond training library
Solution Approach 1:
The patent applies different processing strategies to different frequency regions and audio scenarios rather than using a uniform approach. The neural network incorporates locally adapted parameters and processing methods tailored to specific frequency bands and acoustic conditions, improving performance on unseen scenarios by handling each case with appropriate local expertise.
Solution Approach 2:
The system performs preliminary analysis of the input audio characteristics before applying bandwidth extension, allowing it to adapt its processing approach based on the specific scenario. This preliminary action enables the network to prepare appropriate processing parameters in advance, improving generalization to unseen scenarios by dynamically adjusting to match the actual input characteristics.
Data Source
AI summary
A bandwidth extension (BWE) method includes: determining parameters of a low-frequency spectrum of a narrowband signal; inputting the parameters of the low-frequency spectrum into a neural network model, and obtaining a correlation parameter based on an output of the neural network model; obtaining a target high-frequency amplitude spectrum based on the correlation parameter and a low-frequency amplitude spectrum; obtaining a high-frequency spectrum based on a low-frequency phase spectrum and the target high-frequency amplitude spectrum of the narrowband signal; and obtaining a broadband signal after BWE based on the low-frequency spectrum and the high-frequency spectrum.