Spiking Neural Network Audio Keyword Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current keyword spotting solutions for battery-powered systems face high power consumption, limiting their ability to recognize keywords and being inefficient in noise environments, especially for always-on systems.
Innovation Solution
A spiking neural network-based system that converts audio features into spikes, using a sparsely active network to reduce power consumption by processing data only when necessary, allowing for continuous keyword detection without separate acoustic and language models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional keyword spotting solutions are used in battery-powered systems, then keyword recognition capability is limited, but power consumption is high
Solution Approach 1:
The system dynamically adjusts its processing mode between wake word detection and keyword spotting based on operational state. The spiking neural network processes audio data only when necessary (event-driven), transitioning between active and idle states to optimize the balance between recognition capability and power consumption
Solution Approach 2:
The patent changes the computational parameters by using spiking neural networks with binary spike representations instead of traditional floating-point computations. This parameter change reduces computational complexity and power consumption while maintaining keyword recognition capability in battery-powered systems
2Productivity
If always-on keyword spotting is implemented, then continuous keyword detection is achieved, but power overhead increases significantly
Solution Approach 1:
The system uses periodic wake word detection intervals rather than continuous processing. The spiking neural network remains in a low-power state between wake word triggers, activating full keyword spotting processing only periodically when wake words are detected, thereby reducing power overhead while maintaining continuous detection capability
Solution Approach 2:
The patent extracts and separates the wake word detection function from the full keyword spotting function. By taking out the always-on requirement and implementing it only for wake word detection, the system reduces power overhead while maintaining continuous keyword detection capability when activated
3Measurement precision
If larger model size is used to achieve higher accuracy, then keyword detection accuracy improves, but power consumption increases
Solution Approach 1:
The patent replaces traditional mechanical computing systems with spiking neural networks that use binary spike events instead of continuous numerical computations. This substitution maintains keyword detection accuracy while significantly reducing power consumption by eliminating the need for complex arithmetic operations in larger models
Data Source
AI summary
An example apparatus for detecting keywords in audio includes an audio receiver to receive audio comprising a keyword to be detected. The apparatus also includes a spike transducer to convert the audio into a plurality of spikes. The apparatus further includes a spiking neural network to receive one or more of the spikes and generate a spike corresponding to a detected keyword.