Blind Bandwidth Extension Using K-Means and SVM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Bandwidth extension for music signals poses challenges due to the importance of fine high-band structures, which existing methods like LPC-based approaches may not directly address, requiring more precise predictors.
Innovation Solution
An unsupervised clustering method and supervised regression process are used to generate prediction models, which are applied to extract and enhance subbands in musical audio signals, employing techniques like k-means clustering and support vector machines, to predict and generate high-frequency content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If LPC-based bandwidth extension methods are used for music signals, then the processing complexity is reduced, but the prediction precision of high-frequency components deteriorates
Solution Approach 1:
The patent segments the music signal into multiple subbands using a time-frequency transform, then applies different prediction models to different subbands. This segmentation allows the system to capture fine high-band structures in music signals while maintaining manageable processing complexity through localized analysis.
Solution Approach 2:
The patent dynamically selects prediction models based on the characteristics of each subband using unsupervised clustering. Instead of using a fixed LPC-based approach, the system adapts the prediction method to match the local signal characteristics, improving prediction precision for different types of musical content.
2Quantity of substance
If blind bandwidth extension without side information is used, then the data requirement is reduced, but the reconstruction quality of high-frequency content deteriorates
Solution Approach 1:
The patent performs preliminary unsupervised clustering on the available low-frequency subbands to identify signal characteristics and select appropriate prediction models before generating high-frequency content. This preliminary analysis enables accurate blind prediction without requiring transmitted side information.
Solution Approach 2:
The patent replaces traditional mechanical filtering approaches with machine learning-based prediction models (including support vector machines and neural networks) that can accurately predict high-frequency content from low-frequency inputs without requiring additional transmitted data.
3Ease of manufacture
If spectral band replication is used to generate high-frequency content, then the algorithm simplicity is maintained, but the harmonic structure matching deteriorates
Solution Approach 1:
The patent applies different prediction models to different subbands based on their local characteristics identified through clustering. This allows the system to maintain algorithm simplicity through modular processing while achieving accurate harmonic structure matching by tailoring the prediction approach to each subband's specific properties.
Data Source
AI summary
A system and method of blind bandwidth extension. The system selects a prediction model from a number of stored prediction models that were generated using an unsupervised clustering method (e.g., a k-means method) and a supervised regression process (e.g., a support vector machine), and extends the bandwidth of an input musical audio signal.