Hearing Device Neural Network Switching for Low-Power Sound Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hearing devices face challenges in efficiently enhancing sound quality and intelligibility due to noise interference, particularly in complex environments, with existing deep neural network (DNN) models consuming excessive power and memory resources.
Innovation Solution
Implementing simplified DNN models in hearing devices that adapt to ambient environments by classifying sound categories and selectively loading optimized DNN models and feature extractors, utilizing memory-efficient pruning and quantization techniques to reduce resource consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep neural network models are used to enhance sound quality and intelligibility, then sound enhancement performance is improved, but power consumption and memory resource usage increase excessively
Solution Approach 1:
The patent divides the single large DNN model into multiple specialized sub-models, each trained for specific acoustic environments (e.g., noisy vs. quiet conditions). The system segments the neural network processing based on environmental classification, loading only the appropriate sub-model for the current condition. This segmentation reduces the computational burden and memory requirements compared to running a single comprehensive model, while maintaining sound enhancement quality through specialized processing for each environment type.
Solution Approach 2:
The patent dynamically changes operational parameters by selecting different DNN models based on environmental classification. The system monitors acoustic characteristics and switches between pre-trained models with different parameter sets optimized for specific conditions. This parameter change approach allows the system to maintain high enhancement performance when needed while consuming minimal resources in favorable conditions, directly addressing the power-quality tradeoff.
2Reliability
If deep neural network models are used to enhance sound quality and intelligibility, then sound enhancement performance is improved, but memory resource consumption increases excessively
Solution Approach 1:
The patent segments the neural network model into multiple smaller specialized models, each handling specific acoustic scenarios. By dividing the monolithic DNN into environment-specific sub-models, the system reduces the memory footprint required at any given time. Only the relevant sub-model for the current acoustic environment is loaded into memory, rather than maintaining all possible processing capabilities simultaneously.
Solution Approach 2:
The system dynamically changes the memory configuration by loading different model parameters based on environmental classification. When the acoustic environment changes, the system swaps between pre-trained model sets, each with parameters optimized for specific conditions. This dynamic parameter switching allows efficient memory utilization while maintaining high enhancement performance when specialized processing is required.
3Adaptability or versatility
If multiple DNN models are maintained for different environments, then adaptability to various noise conditions is improved, but device complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training multiple DNN models offline for different acoustic environments before deployment. The classification system and multiple specialized models are prepared in advance, allowing the device to simply select from pre-configured options during operation rather than performing complex real-time model training or switching logic. This preliminary preparation reduces runtime complexity while maintaining high adaptability.
Solution Approach 2:
The system employs feedback mechanisms where the acoustic environment is continuously classified and this classification feedback drives the selection of appropriate pre-trained models. The feedback loop monitors acoustic characteristics and dynamically switches between model sets, providing adaptive performance without requiring complex real-time model management. The feedback-based selection simplifies the overall system architecture compared to alternative approaches requiring intricate model orchestration.
Data Source
AI summary
A persistent memory of an ear-wearable device stores a plurality of neural network data objects each defining a respective neural network. The ear-wearable device includes a digital signal processor comprising a neural network processor. The digital signal processor is operable to: classify an ambient environment of a sound signal into one of a plurality of classifications; select one of the neural network data objects to enhance the sound signal based on the classification; and load neural network data from the selected neural network data objects into a memory. The neural network processor enhances the sound signal using the neural network data. The ear-wearable device includes an audio processing circuit that reproduces the enhanced sound signal via a receiver of the ear-wearable device.