Ear-Worn Neural Spatial Focusing Against Interfering Speakers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional ear-worn devices face challenges in reducing noise, particularly from interfering speakers, due to limitations in beamforming patterns, which are distorted by the wearer's anatomy and perform better on high-frequency sounds, and struggle in reverberant environments, with limited sound reduction capabilities and noise addition in quiet settings.
Innovation Solution
Implementing neural networks for spatial focusing in ear-worn devices, using multiple microphones to apply different weights based on sound direction, distinguishing target and interfering sounds, and independently controlling background noise and speech volumes to enhance target speech.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional beamforming patterns are used to reduce noise from interfering speakers, then sound reduction capability is improved, but the beamforming pattern becomes warped due to interference from the wearer's head, torso, and ear
Solution Approach 1:
The patent applies neural network-based spatial filtering that dynamically adapts to the wearer's anatomy by learning the transfer functions between microphones and the ear canal. This changes the approach from fixed geometric beamforming patterns to adaptive filtering that compensates for anatomical interference, resolving the contradiction between noise reduction and pattern accuracy
Solution Approach 2:
The patent replaces the mechanical/physical beamforming approach (which relies on fixed microphone array geometry and signal delays) with a neural network-based computational approach. This substitution allows the system to learn and adapt to individual anatomical variations, eliminating the warping effect caused by head, torso, and ear interference
2Measurement precision
If conventional beamforming is used to focus on sounds from the front, then intelligibility of target speech is improved, but sounds from reverberant paths entering from the front are not attenuated
Solution Approach 1:
The neural network spatial filter learns from training data that includes reverberant environments, using feedback mechanisms to adaptively adjust filtering parameters. The system continuously optimizes its transfer function estimates to account for reverberation paths, enabling it to attenuate reverberant noise while preserving direct speech paths
Solution Approach 2:
The patent changes the filtering parameters dynamically based on the acoustic environment. The neural network adapts its transfer function estimates in real-time to distinguish between direct speech paths and reverberant paths, allowing selective attenuation of reverberant noise while maintaining speech intelligibility
3Measurement precision
If conventional beamforming patterns are applied, then high-frequency sound localization is improved, but low-frequency sound reduction is limited
Solution Approach 1:
The patent implements a dynamic neural network-based spatial filter that adapts its characteristics across different frequency ranges. The system dynamically adjusts its filtering parameters based on the frequency content of the input signal, enabling effective low-frequency noise reduction while maintaining accurate sound localization through learned transfer functions that account for frequency-dependent anatomical effects
4Object-affected harmful factors
If neural networks are used for spatial focusing to reduce interfering speaker sound, then noise reduction capability is improved, but device complexity increases
Solution Approach 1:
The patent performs preliminary training of the neural network offline using recorded data from individual wearers. This preliminary action creates pre-computed transfer function estimates that can be applied in real-time without requiring complex computational resources during actual use. The heavy computational lifting is done in advance, simplifying the deployed device complexity while maintaining high noise reduction capability
Data Source
AI summary
An ear-worn device includes two or more microphones and noise reduction circuitry including neural network circuitry. The neural network circuitry is configured to: receive multiple audio signals wherein at least two of the multiple audio signals each originate from a different one of the two or more microphones and/or at least one of the multiple audio signals is a beamformed audio signal originating from the two or more microphones; and implement one or more neural network layers trained to perform background noise modification and spatial focusing based on the multiple audio signals, such that the neural network circuitry generates, based on the multiple audio signals, one or more neural network outputs. The noise reduction circuitry is configured to output, based on the one or more neural network outputs, an output audio signal comprising a background noise-modified and spatially-focused version of a first audio signal of the multiple audio signals.