Binary Neural Network Hearing Aid for Speech-Noise Separation

Solution Overview

Problem

Existing hearing devices with neural networks require substantial hardware and energy due to large, complex networks for distinguishing speech components from noise, which is undesirable for compact and energy-efficient designs.

Innovation Solution

Implementing a neural network with binary weights and values in hearing devices, using XNOR operations and sign-based activation functions to reduce memory and computational requirements, allowing for compact and energy-efficient operation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large neural networks with many neurons are used to distinguish speech components from noise, then the accuracy of speech-noise distinction is improved, but the hardware resources and energy consumption increase substantially

Engineering Contradiction:

Improvespeech-noise distinction accuracyVSAvoidenergy consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies parameter changes by transforming the neural network weights and activations from continuous floating-point values to discrete binary values (-1 and +1). This quantization dramatically reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby reducing energy consumption while preserving sufficient accuracy for speech-noise distinction in hearing devices.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the traditional mechanical/computational approach of using large-scale floating-point neural networks with a binary neural network architecture that uses bitwise operations. This substitution replaces complex arithmetic operations with simpler logical operations (XNOR and population count), significantly reducing hardware requirements and energy consumption while maintaining functional equivalence for audio signal processing.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If large neural networks with many neurons are used to distinguish speech components from noise, then the accuracy of speech-noise distinction is improved, but the hardware resources increase substantially

Engineering Contradiction:

Improvespeech-noise distinction accuracyVSAvoidhardware resources

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by transforming the neural network weights and activations from continuous floating-point values to discrete binary values (-1 and +1). This quantization dramatically reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby reducing energy consumption while preserving sufficient accuracy for speech-noise distinction in hearing devices.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the traditional mechanical/computational approach of using large-scale floating-point neural networks with a binary neural network architecture that uses bitwise operations. This substitution replaces complex arithmetic operations with simpler logical operations (XNOR and population count), significantly reducing hardware requirements and energy consumption while maintaining functional equivalence for audio signal processing.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Use of energy by moving object

If binary weights and values are used in the neural network, then hardware resources and energy consumption are reduced, but the computational accuracy may be compromised

Engineering Contradiction:

Improveenergy consumptionVSAvoidcomputational accuracy

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent employs binary values as simplified, low-cost representations of neural network weights and activations. Although binary representations are less precise than floating-point values, the system compensates through architectural design and training techniques, achieving sufficient accuracy for hearing device applications while dramatically reducing computational complexity and energy consumption.

Inventive Principle:

Principle #27Cheap short-living objects (Disposable)

Solution Approach 2:

The patent applies parameter changes by transforming the neural network weights and activations from continuous floating-point values to discrete binary values (-1 and +1). This quantization dramatically reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby reducing energy consumption while preserving sufficient accuracy for speech-noise distinction in hearing devices.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260019758A1Method for operating a hearing device, and hearing device

Publication Date: 2026.01.15 SIVANTOS PTE LTD

AI summary

A method for operating a hearing device which has a neural network having a plurality of neurons to which a weighting vector with binary weights is in each case assigned. An input vector with binary values is fed to each neuron and is processed with the weighting vector in order to obtain a transfer function. The transfer function is processed with an activation function in such a way that a binary result is provided. There is also described a method for training the neural network, and a hearing device that is configured for carrying out the methods.