Steering Vector Generation for Stable Neural Network Beamforming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing beamformer techniques face numerical instability during neural network training due to eigenvalue decomposition problems and high noise and reverberation levels, leading to inaccurate target sound extraction.
Innovation Solution
A target sound extraction technique that generates a steering vector using a power method to approximate the eigenvector corresponding to the maximum eigenvalue, avoiding the need for eigenvalue decomposition and stabilizing the neural network training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If eigenvalue decomposition is used to determine the steering vector, then the beamformer estimation can be performed, but the calculation becomes numerically unstable during neural network training
Solution Approach 1:
The patent extracts only the essential component (the eigenvector corresponding to the maximum eigenvalue) needed for beamformer estimation, rather than performing complete eigenvalue decomposition. This is achieved by using the power method to iteratively compute only the dominant eigenvector, eliminating the numerical instability of full eigenvalue decomposition while maintaining the necessary functionality.
Solution Approach 2:
The patent changes the computational approach from exact eigenvalue decomposition to an approximate method (power method) that iteratively converges to the dominant eigenvector. This parameter change in the calculation method transforms an unstable operation into a stable iterative process that is compatible with neural network training.
2Ease of manufacture
If inverse matrix operation is used to estimate the beamformer, then the calculation is simpler, but the approximation error increases in noisy environments
Solution Approach 1:
The patent introduces the power method as an intermediary computational approach between exact eigenvalue decomposition and simple inverse matrix operation. This intermediary method provides a balanced solution that maintains better accuracy than inverse matrix operations in noisy environments while avoiding the complexity and instability of full eigenvalue decomposition.
Data Source
AI summary
Provided is a target sound extraction technique based on a steering vector generation method enabling instability in a calculation to be prevented when a neural network is trained by using an error back propagation method to reduce an estimation error of a beamformer. A target sound signal generation apparatus generates a target sound signal yt,f corresponding to a target sound included in an observed sound from an observed signal vector xt,f corresponding to the observed sound collected by using a plurality of microphones. The target sound signal generation apparatus includes a mask generation unit, a steering vector generation unit, a beamformer vector generation unit, and a target sound signal generation unit. The mask generation unit is configured as a neural network trained by using an error back propagation method. The steering vector generation unit generates a steering vector hf by determining an eigenvector corresponding to a maximum eigenvalue of a predetermined matrix generated from the observed signal vector xt,f and a mask γt,f by using a power method.