Sound Recognition with Random FFT Numbers and Hop Lengths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sound recognition technologies face challenges in accurately identifying and distinguishing different types of sounds due to variations in data resolution and hop lengths, leading to inconsistent performance in neural network models.
Innovation Solution
The method involves generating multiple two-dimensional feature maps using random Fast Fourier Transform (FFT) numbers and hop lengths to create diverse training data for neural networks, enhancing sound recognition by varying frequency and time axis resolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed FFT numbers and hop lengths are used for sound recognition, then the processing method is simple, but the recognition accuracy is insufficient due to inconsistent performance across different sound types
Solution Approach 1:
The patent applies dynamics by making the FFT number and hop length variable rather than fixed. The system dynamically adjusts these parameters based on the characteristics of the input sound signal, allowing the processing method to adapt to different sound types and improve recognition accuracy while managing complexity through conditional logic.
Solution Approach 2:
The patent directly implements parameter changes by varying the FFT number and hop length according to sound characteristics. Different parameter sets are selected based on the detected sound type, enabling the system to optimize processing for each specific sound category and resolve the contradiction between simple processing and accurate recognition.
2Reliability
If multiple 2D feature maps with random FFT numbers and hop lengths are generated, then sound recognition accuracy is improved, but the data processing complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-defining multiple sets of FFT numbers and hop lengths that can be selected based on sound characteristics. This preparation work is done in advance, allowing the system to quickly switch between parameter sets during processing without real-time complex calculations, thus improving reliability while controlling processing complexity.
Solution Approach 2:
The patent segments the processing into distinct stages: first detecting sound characteristics, then selecting appropriate parameter sets, and finally generating feature maps with those parameters. This segmentation allows the system to handle multiple feature maps efficiently by processing them in organized batches rather than simultaneously, reducing overall complexity.
3Measurement precision
If varied FFT parameters are used to create diverse training data, then neural network training effectiveness is improved, but the computational resources required increase
Solution Approach 1:
The patent applies local quality by applying different FFT parameters selectively based on local characteristics of the sound signal. Rather than using maximum complexity parameters for all data, the system chooses appropriate parameter sets for each sound segment, achieving good training accuracy while reducing unnecessary computational energy consumption on simpler sound portions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An example sound recognition method may include sampling input sound based on a preset sampling rate; performing Fast Fourier Transform (FFT) on the sampled input sound based on at least one of random FFT numbers or random hop lengths, and generating a two-dimensional (2D) feature map with a time axis and a frequency axis from the sampled input sound on which FFT is performed; training a neural network model, which recognizes sound, with a plurality of 2D feature maps including the first 2Dfeature map and an nth 2D feature map as training data.