Audio Separation Using Convolution Layers With Different Kernel Lengths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current audio separation methods, such as Conv-TasNet, suffer from lower frequency resolution due to identical and relatively short one-dimensional convolution kernel lengths, leading to inadequate separation of audio mixture signals, particularly in resolving multiple time-frequency regions between sound harmonics.
Innovation Solution
The proposed audio processing method employs convolution layers and deconvolution layers with different kernel lengths to improve the neural network architecture's performance, allowing for better analysis and synthesis of audio signals by using multi-layer one-dimensional convolution kernels in both layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If identical and relatively short one-dimensional convolution kernel lengths are used in convolution layers, then the device complexity is reduced and ease of manufacture is improved, but the frequency resolution deteriorates and manufacturing precision worsens
Solution Approach 1:
The patent segments the audio processing task by using multiple convolution layers with different kernel lengths (first, second, and third kernel lengths) instead of a single uniform kernel length. This segmentation allows each layer to capture different temporal patterns and frequency components, thereby improving frequency resolution while maintaining implementation feasibility through modular architecture.
Solution Approach 2:
The patent applies local quality by assigning different kernel lengths to different convolution layers based on their specific processing needs. The first convolution layer uses a first kernel length, the second uses a second kernel length, and the third uses a third kernel length, allowing each layer to have optimized local characteristics for its specific function in the audio separation pipeline.
2Measurement precision
If multiple convolution layers with different kernel lengths are used, then the frequency resolution and audio separation performance are improved, but the device complexity increases
Solution Approach 1:
The patent introduces dynamic characteristics by varying the kernel lengths across different convolution layers rather than using a static uniform length. This dynamic approach allows the network to adaptively capture both short-term and long-term temporal dependencies in audio signals, improving frequency resolution while the systematic arrangement keeps complexity manageable.
Solution Approach 2:
The patent achieves multi-functionality by designing a unified convolutional neural network architecture that performs multiple functions through its layered structure. The same basic convolution layer template is reused with different kernel lengths to achieve both temporal feature extraction and frequency resolution enhancement, reducing overall system complexity despite the varied kernel lengths.
3Ease of manufacture
If traditional Conv-TasNet with uniform short kernel lengths is used, then the model training and implementation are simpler, but the separation quality of audio mixture signals deteriorates
Solution Approach 1:
The patent directly applies parameter changes by modifying the kernel length parameter across different convolution layers. Instead of using uniform short kernel lengths as in traditional Conv-TasNet, the patent varies the kernel length parameter (first, second, and third lengths) to improve frequency resolution and separation quality while maintaining the overall model structure for manageable implementation complexity.
Data Source
AI summary
The audio processing method may include: utilizing a first convolution layer circuit and a second convolution layer circuit to perform convolution processing on an audio mixture signal, respectively, in order to generate an input feature map; performing separation processing corresponding to a predetermined neural network model on the input feature map to obtain a first output feature map and a second output feature map respectively corresponding to a first type and a second type; and utilizing a first deconvolution layer circuit and a second deconvolution layer circuit to perform deconvolution processing on different portions of the first output feature map, respectively, and utilizing the first deconvolution layer circuit and the second deconvolution layer circuit to perform deconvolution processing on different portions of the second output feature map, respectively, in order to generate corresponding processing results, for generating multiple audio output signals.