Neural Beamforming From Phase Difference for Noisy Speech Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing beamforming techniques require significant computational resources and struggle to accurately determine spatial information in noisy environments, particularly in scenarios with moving speakers or close microphone placements.
Innovation Solution
A beamforming method utilizing a neural network that processes phase differences between sound signals from two microphones, employing element-wise multiplication and gain control, without the need for computing steering vectors or spatial correlation matrices, and incorporating a lightweight auto-encoder for miniaturization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule-based beamforming is used to obtain spatial information, then beamforming can be implemented, but a large amount of computation is required
Solution Approach 1:
The patent replaces the mechanical computation system (rule-based beamforming algorithms requiring steering vectors and spatial correlation matrices) with a neural network system that learns spatial information automatically from phase differences, eliminating the need for complex mathematical computations
Solution Approach 2:
The patent changes the input parameters from complex spatial information (steering vectors, correlation matrices) to simple phase difference values between microphones, which the neural network processes to achieve beamforming with reduced computational requirements
2Measurement precision
If conventional beamforming techniques are used, then spatial information can be obtained, but the computation amount is large and accuracy is reduced in noisy environments
Solution Approach 1:
The patent extracts only the essential phase difference information from the microphone signals, discarding the computationally intensive steering vectors and spatial correlation matrices, while retaining the ability to achieve accurate beamforming through neural network processing
Solution Approach 2:
The patent introduces a neural network as an intermediary between the simple phase difference measurements and the final beamforming output, enabling the system to achieve high accuracy in noisy environments without requiring complex conventional beamforming computations
3Device complexity
If neural network is used for beamforming, then computation amount is reduced, but the neural network structure must be optimized for beamforming
Solution Approach 1:
The patent defines specific parameter constraints for the neural network including input phase difference values between -π and π, and output mask values between 0 and 1, which guides the network design and simplifies implementation while maintaining computational efficiency
Solution Approach 2:
The patent employs training procedures with feedback mechanisms where the neural network learns optimal beamforming parameters from training data, automatically optimizing the network structure for beamforming applications without manual design complexity
Data Source
AI summary
Provided are a beamforming method and a beamforming system. The beamforming method may comprise the steps of: receiving a first sound signal and a second sound signal by respectively using a first microphone and a second microphone spaced apart from the first microphone by a pre-determined distance; obtaining a Fourier transform result for each of the first sound signal and the second sound signal; obtaining a phase difference between the first sound signal and the second sound signal from the Fourier transform result; performing an arithmetic operation by inputting the phase difference to a beamforming model by using a neural network; performing element multiplication on an operation result of the neural processor and the Fourier transform result for the first sound signal; and outputting a result of the element multiplication.