Binary Neural Network Hearing Aid for Speech-Noise Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for distinguishing speech from background noise in hearing aids require large hardware and energy resources, which are not suitable for the compact and energy-efficient design of hearing aids.
Innovation Solution
Implementing a neural network with binary weights and values, using XNOR operations and sign-based activation functions, reduces hardware and energy requirements by representing weights and values with a single bit, allowing efficient processing and prediction of audio signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large neural networks with many layers are used to distinguish speech from background noise, then speech intelligibility is improved, but hardware resources and energy consumption increase significantly
Solution Approach 1:
The patent applies parameter changes by transforming the neural network weights and activations from multi-bit floating-point representations to binary representations. This fundamental parameter change reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby significantly reducing energy consumption while preserving speech intelligibility through careful quantization and training strategies
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 bit-level operations. This substitution replaces energy-intensive arithmetic operations with efficient bitwise operations (XNOR and population count), achieving the same speech enhancement function with dramatically reduced hardware resources and energy consumption
2Measurement precision
If large neural networks with many layers are used to distinguish speech from background noise, then speech intelligibility is improved, but hardware size increases
Solution Approach 1:
The patent applies parameter changes by transforming the neural network weights and activations from multi-bit floating-point representations to binary representations. This fundamental parameter change reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby significantly reducing energy consumption while preserving speech intelligibility through careful quantization and training strategies
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 bit-level operations. This substitution replaces energy-intensive arithmetic operations with efficient bitwise operations (XNOR and population count), achieving the same speech enhancement function with dramatically reduced hardware resources and energy consumption
3Use of energy by moving object
If binary weights and values are used with XNOR operations, then hardware resources and energy consumption are reduced, but computational accuracy may be compromised
Solution Approach 1:
The patent employs copying by training the binary neural network on large datasets to learn optimal binary weight patterns that replicate the functionality of full-precision networks. The binary network is trained to copy the speech enhancement behavior of conventional neural networks, achieving comparable computational accuracy through statistical learning rather than exact arithmetic computation
Solution Approach 2:
The patent applies parameter changes by transforming the neural network weights and activations from multi-bit floating-point representations to binary representations. This fundamental parameter change reduces the computational complexity from multiply-accumulate operations to simpler XNOR and population count operations, thereby significantly reducing energy consumption while preserving speech intelligibility through careful quantization and training strategies
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The invention relates to a method (28) for operating a hearing aid (2) comprising a neural network (18) with several neurons (22), each of which is assigned a weighting vector (40) with binary weights (44). Each neuron (22) is supplied with an input vector (34) with binary values (36) and processed with the weighting vector (40) to obtain a transfer function (42). The transfer function (42) is processed with an activation function (48) such that a binary result (26) is provided. The invention further relates to a method (52) for training a neural network (18) and a hearing aid (2).