Hybrid Auditory Filterbanks for Stable Adaptive Audio Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio processing techniques face challenges in achieving a balance between interpretability and adaptability, particularly in encoder-mask-decoder settings, with fixed transforms being inflexible and data-driven methods being unstable and hard to interpret.
Innovation Solution
A hybrid auditory filterbank is introduced, combining fixed and trainable filters to enhance stability and adaptability, using an auditory filterbank structure with trainable convolutional layers and a penalty term to maintain tightness, ensuring numerical stability and efficient reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If fixed time-frequency transforms (STFT, CQT) are used as encoders, then interpretability and control are improved, but adaptability to different tasks deteriorates
Solution Approach 1:
The patent merges fixed time-frequency transforms with trainable neural network components to create a hybrid encoder architecture. The fixed transform provides interpretable frequency decomposition while trainable layers adapt to specific tasks, combining the advantages of both fixed and data-driven approaches.
Solution Approach 2:
The patent introduces dynamic adaptability by making parts of the encoder trainable while keeping other parts fixed. This allows the system to adapt its parameters dynamically during training for different tasks while maintaining the structured interpretability of the fixed transform components.
2Adaptability or versatility
If data-driven feature extraction methods are used, then flexibility and adaptability are improved, but stability and interpretability deteriorate
Solution Approach 1:
The patent combines data-driven trainable components with fixed, stable transform operations. The trainable parts provide flexibility and adaptability while the fixed transform components ensure numerical stability and interpretability, resolving the contradiction between flexibility and stability.
3Adaptability or versatility
If purely trainable filterbanks are used, then adaptability to tasks is improved, but numerical stability and reconstruction quality deteriorate
Solution Approach 1:
The patent applies different properties to different parts of the filterbank: some filters are fully trainable for task optimization while others are fixed to ensure numerical stability. This local differentiation allows simultaneous achievement of adaptability and stability.
Solution Approach 2:
The filterbank is constructed as a composite structure combining trainable and fixed components, analogous to composite materials. This hybrid construction allows the system to benefit from both the adaptability of trainable filters and the stability of fixed transforms.
Data Source
Figure 1
Figure 2~4
Figure 3
AI summary
Disclosed is a hybrid auditory filterbank (100) configured for audio processing implemented on a data processing apparatus. The hybrid auditory filterbank (100) comprises a plurality of filters (102) configured to decompose an input audio signal (104) into a plurality of sub-bands (106). The filters (102) are each based on a fixed filter (108) and a trainable filter (110). The fixed filters (108) each comprise a filter of an auditory filterbank. The trainable filters (110) have been trained to perform an audio processing task and to improve the stability of the hybrid auditory filterbank (100).