Keyword Spotting Neural Network Depthwise Pointwise Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional keyword spotting systems face a trade-off between high accuracy and low latency, requiring significant computational resources, which is particularly challenging on mobile devices where real-time performance is essential.
Innovation Solution
The proposed solution involves a keyword spotting method using a temporal convolutional neural network (TC-ResNet) that applies 1D convolution along the temporal dimension, treating mel-frequency cepstral coefficients (MFCC) as input channels, reducing the number of operations and enhancing accuracy while maintaining a compact architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional convolution-based keyword spotting is used, then accuracy is improved, but computational complexity and latency increase significantly
Solution Approach 1:
The patent segments the convolution operation into depthwise convolution and pointwise convolution stages. The depthwise convolution applies separate filters to each input channel, and the pointwise convolution combines the results using 1x1 convolutions. This segmentation reduces the total number of computations while maintaining accuracy, directly resolving the contradiction between accuracy and computational complexity.
Solution Approach 2:
The patent changes the dimensionality of the convolution operation by applying filters along the depth dimension rather than treating all channels uniformly. This dimensional approach allows for more efficient computation by processing channels independently and then combining results, reducing computational complexity while preserving the accuracy needed for keyword spotting.
2Measurement precision
If traditional convolution-based keyword spotting is used, then accuracy is improved, but processing speed decreases
Solution Approach 1:
By segmenting the convolution into depthwise and pointwise operations, the patent enables more efficient parallel processing. The depthwise convolution processes each channel independently, and the pointwise convolution combines results efficiently, significantly improving processing speed while maintaining the accuracy required for keyword spotting.
3Measurement precision
If traditional convolution-based keyword spotting is used, then accuracy is improved, but resource consumption increases
Solution Approach 1:
The segmentation of convolution operations reduces the total computational load, directly decreasing energy consumption and resource usage. By processing channels independently in the depthwise stage and combining results efficiently in the pointwise stage, the patent maintains accuracy while significantly reducing the computational resources required.
Data Source
AI summary
A keyword spotting apparatus, method, and computer-readable recording medium are disclosed. The keyword spotting method using an artificial neural network according to an embodiment of the disclosure may include obtaining an input feature map from an input voice; performing a first convolution operation on the input feature map for each of n different filters having the same channel length as the input feature map, wherein a width of each of the filters is w1 and the width w1 is less than a width of the input feature map; performing a second convolution operation on a result of the first convolution operation for each of different filters having the same channel length as the input feature map; storing a result of the second convolution operation as an output feature map; and extracting a voice keyword by applying the output feature map to a learned machine learning model.