Dynamic Neural Network Hearing Devices for Power-Aware Sound Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hearing devices face challenges in efficiently utilizing deep neural networks (DNNs) for sound enhancement due to limited processing capability, power consumption, and resource constraints, hindering widespread adoption.
Innovation Solution
Implementing a dynamically variable DNN that adjusts its state in response to audio and system changes, such as sparsity, numerical precision, processing frequency, and latent representation, to optimize resource utilization and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep neural network is used for sound enhancement in hearing devices, then sound processing quality is improved, but power consumption and resource usage increase
Solution Approach 1:
The neural network operates in multiple dynamic states (low-power state and high-performance state) that can be switched based on processing needs. The system dynamically adjusts the operational mode of the DNN to balance between power consumption and processing quality, transitioning to low-power state during normal operation and high-performance state when enhanced sound processing is required.
Solution Approach 2:
The system changes key parameters of the neural network operation including precision levels, activation thresholds, and processing frequency. By adjusting these parameters dynamically, the system can reduce computational load and power consumption while maintaining adequate sound enhancement quality under varying listening conditions.
2Measurement precision
If a deep neural network is used for sound enhancement, then inference accuracy is improved, but device complexity increases
Solution Approach 1:
The neural network processing is segmented into different operational modes and precision levels. The system divides the DNN into core operations that always execute and optional operations that execute conditionally, allowing complex inference capabilities to be distributed across different processing layers and executed only when needed.
Solution Approach 2:
The system implements partial neural network processing by selectively activating only the necessary portions of the DNN based on current processing requirements. Instead of always running the full complex network, the system applies partial computation with reduced precision or skipped layers when full accuracy is not required, thereby reducing device complexity demands.
3Measurement precision
If the DNN operates at high precision, then sound enhancement quality is improved, but resource consumption increases
Solution Approach 1:
Different precision levels are applied to different parts of the neural network and different processing scenarios. Critical path operations and sensitive computations use higher precision, while less critical operations use lower precision. This localized quality approach maintains sound enhancement quality where needed while reducing overall resource consumption.
Solution Approach 2:
The system dynamically changes numerical precision parameters based on operational context. Precision can be adjusted between operations, data types, and computational stages, allowing the system to use lower precision (and thus fewer resources) when high accuracy is not critical, while maintaining high precision when sound enhancement quality requires it.
Data Source
AI summary
An ear-wearable device includes a digital signal processor that receives an audio signal from sound source of the ear-wearable device and reproduces the audio signal at a receiver that is placed within an ear of a user. The digital signal processor includes a deep neural network (DNN) logic circuit. The DNN logic circuit is operable to perform sound enhancement on the audio signal using a DNN stored in memory. An audio feature detector of the ear-wearable device is operable to detect an audio change via the digital signal processor that triggers a change of state of the DNN, the change of state affecting resource consumption by the DNN. The audio feature detector applies the change of state to the DNN. The DNN performs the sound enhancement in the changed state.