Adaptive Sampling for Sound Recognition Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional sound recognition model training methods face issues with high costs and reduced precision due to repetitive and duplicated training audio files, caused by fixed-length sampling methods that either repeat or duplicate sound events to match sampling units, leading to overemphasis on specific classifications and altered characteristic distributions.
Innovation Solution
A method that determines a relationship between sound events and parameters to generate training audio files with varying lengths and time differences, using a processing device to sample and preprocess audio data, thereby creating a sound recognition model for accurate classification without the need for large-scale deep learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Duration of action of moving object
If fixed-length sampling with duplication is used to compensate for insufficient sound event duration, then training audio files can be generated to meet length requirements, but the training files contain many identical and continuous sound events which changes the original characteristic distribution and decreases recognition precision
Solution Approach 1:
The patent applies dynamics by making the sampling parameters adaptive rather than fixed. The sampling shift and sampling length are dynamically adjusted based on the actual sound event duration and characteristics. For sound events shorter than the sampling unit, instead of simple duplication, the system calculates an adaptive sampling shift that distributes samples across different time positions, and adjusts sampling length to match the actual event duration, thereby maintaining characteristic distribution while meeting length requirements.
Solution Approach 2:
The patent changes the parameters of sampling (sampling shift and sampling length) based on the relationship between sound event duration and sampling unit length. When sound event duration is less than sampling unit length, the system modifies the sampling shift parameter to be proportional to the event duration, and adjusts sampling length accordingly. This parameter adaptation prevents identical continuous samples and preserves the original characteristic distribution.
2Productivity
If conventional fixed-length sampling is used for sound events longer than sampling unit length, then training audio files are generated, but the sampling produces too many repetitive files causing the recognition model to be too concentrated on specific sound classifications
Solution Approach 1:
The patent applies dynamics by making the sampling shift adaptive rather than fixed. For sound events longer than the sampling unit, the system calculates an adaptive sampling shift based on the ratio of event duration to sampling unit length. This dynamic adjustment ensures that samples are distributed across different time positions rather than being repetitive, allowing the model to learn from diverse temporal patterns and improve adaptability across different sound classifications.
Solution Approach 2:
The patent applies segmentation by dividing long sound events into multiple sampling units with adaptive shifts. Instead of simply overlapping fixed-length windows that produce repetitive samples, the system segments the event into distinct sampling positions determined by the adaptive shift calculation. This creates diverse training samples from single long events, improving both productivity and classification coverage.
3Measurement precision
If large-scale deep learning with detailed time-related characteristics is used for sound recognition, then recognition capability is improved, but training cost and complexity increase significantly
Solution Approach 1:
The patent applies preliminary action by performing adaptive sampling and preprocessing of sound events before model training. The system pre-calculates optimal sampling parameters based on event duration characteristics, generates diverse training samples with appropriate length and temporal distribution, and prepares the data in advance. This preliminary preparation reduces the need for complex large-scale deep learning architectures while maintaining recognition capability, thereby reducing training complexity.
Data Source
AI summary
A sound recognition model training method comprises determining a relationship between a sound event and first parameter and deciding a second parameter in response to the relationship, performing sampling on the sound event using the first parameter and the second parameter to generate training audio files, and inputting at least part of the training audio files to a sound recognition model for training the sound recognition model, wherein a length of each of the training audio files is associated with the first parameter, a time difference between every two of the training audio files is associated with the second parameter, and the sound recognition model is used for determining a sound classification.