Source Separation via Time-Depth Conversion and ASPP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional music source separation methods using Deep Neural Networks face issues such as large model size, significant delay, and poor computation efficiency due to direct processing of mixed signals in the time domain, which do not adequately consider phase information.
Innovation Solution
The method employs space-depth conversion instead of traditional upsampling and incorporates a receptive field-preserving decoder, utilizing Atrous Spatial Pyramid Pooling (ASPP) and temporal attention with dynamic convolution kernels to improve separation efficiency and reduce model parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If direct processing of mixed signals in the time domain is used, then phase information is considered, but model size becomes large and computation efficiency deteriorates
Solution Approach 1:
The patent transforms the time-domain signal processing problem into a depth-domain problem by applying space-depth conversion. Instead of processing time-domain signals directly with large models, the method converts temporal information into spatial depth dimensions, allowing the use of lighter-weight convolutional neural networks that operate on the transformed representation, thereby maintaining phase information while reducing computational burden.
Solution Approach 2:
The patent changes the representation parameters of the signal by converting from time-domain to depth-domain through space-depth transformation. This parameter change allows the system to work with transformed features rather than raw time-domain signals, enabling efficient processing with reduced model complexity while preserving essential phase characteristics.
2Measurement precision
If traditional upsampling is used in decoding, then signal resolution is improved, but receptive field is lost and model complexity increases
Solution Approach 1:
The patent inverts the conventional approach by applying space-depth conversion during decoding instead of traditional upsampling. Rather than increasing temporal resolution through upsampling operations that expand model complexity, the method converts depth information back to time-domain resolution, achieving high signal resolution while maintaining a simpler model architecture through the inverse transformation.
Solution Approach 2:
The patent uses depth-to-time conversion as the inverse operation of space-depth conversion. This dimensional transformation allows the decoder to restore temporal resolution without requiring complex upsampling operations, thereby achieving high signal resolution while keeping the model structure simple and efficient.
3Reliability
If large number of parameters are used to achieve desired separation effect, then separation performance is improved, but model size becomes excessive and delay increases
Solution Approach 1:
The patent replaces the conventional time-domain signal processing mechanism with a depth-domain transformation approach. By substituting direct time-domain processing with space-depth conversion followed by depth-to-time conversion, the system achieves effective source separation with fewer parameters, reducing model size while maintaining separation performance through the transformed representation.
Data Source
AI summary
A source separation method, an apparatus, and a non-transitory computer-readable medium are provided. Atrous Spatial Pyramid Pooling (ASPP) is used to reduce the number of parameters of a model and speed up computation. Conventional upsampling is replaced with a conversion between time and depth, and a receptive field preserving decoder is provided. In addition, temporal attention with dynamic convolution kernel is added, to further achieve lightweight and improve the effect of separation.