Audio Bandwidth Extension Using Aliasing Noise Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional bandwidth extension techniques fail to create physically correct high-frequency components in sound processing.
Innovation Solution
A sound processing method that utilizes a trained model to generate aliasing noise for frequencies higher than the Nyquist frequency, which is then mixed with the original sound signal to produce a fourth sound signal with physically correct high-frequency components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional bandwidth extension techniques are used, then high-frequency signal components can be generated, but the generated high-frequency components are not physically correct
Solution Approach 1:
The patent converts aliasing noise, which is typically considered a harmful artifact of downsampling, into a beneficial resource for generating physically correct high-frequency components. By training a neural network to predict aliasing noise from downsampled signals, the system recovers authentic high-frequency information that would otherwise be lost, transforming a distortion problem into a solution for bandwidth extension.
Solution Approach 2:
The patent applies downsampling to the high-resolution reference signal before feeding it to the neural network, intentionally creating aliasing noise in advance. This preliminary action allows the network to learn the relationship between aliasing patterns and corresponding high-frequency components, enabling it to predict and reconstruct accurate high-frequency content from low-resolution input signals.
2Measurement precision
If a trained model is used to generate aliasing noise, then physically correct high-frequency components can be created, but computational complexity increases
Solution Approach 1:
The patent uses a neural network to copy or replicate the statistical patterns of aliasing noise that would naturally occur in physically correct high-frequency signals. Instead of performing complex physical simulations or requiring high-resolution input, the network learns from training data to generate realistic high-frequency components that mimic the characteristics of genuine aliasing noise, achieving accurate results with computationally efficient operations.
Data Source
AI summary
A sound processing method includes receiving, as an input, a first sound signal sampled at a first sampling frequency. The sound processing method also includes generating, as an output, a second sound signal that is based on aliasing noise for the first sound signal from a frequency range that is higher than a first Nyquist frequency of the first sound signal, using a trained model, in order to produce a third sound signal with a frequency component higher than the first Nyquist frequency. The sound processing method also includes mixing the first sound signal and the third sound signal to create a fourth sound signal.