Prototypical Sound Event Detection With Clustered Negative Prototypes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Sound event detection (SED) is challenging due to the diversity of acoustic events and the need for large amounts of labeled data, which is costly and time-consuming to obtain. Existing models struggle to differentiate between target and non-target sounds and fail to detect rare or unseen events effectively.
Innovation Solution
The method involves training a prototypical network using positive and negative audio samples. Positive prototypes are determined based on positive samples, while negative prototypes are constructed from non-target sound segments within the support set. The network generates an output signal indicating whether a sample belongs to a specific sound event class by comparing the sample with both positive and negative prototypes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large amounts of labeled data are used for training, then detection accuracy improves, but data acquisition cost and time increase
Solution Approach 1:
The patent performs preliminary clustering of negative samples into multiple groups before training, creating organized negative prototypes in advance. This preprocessing step reduces the complexity of training with large unlabeled datasets, enabling accurate detection without requiring extensive manual labeling time.
Solution Approach 2:
The patent creates prototype representations (positive and negative) that serve as simplified copies of actual sound events. These prototypes capture the essential characteristics of sound classes without requiring the full original datasets, reducing data acquisition needs while maintaining detection accuracy.
2Device complexity
If traditional prototypical networks are used, then model simplicity is maintained, but ability to differentiate target from non-target sounds deteriorates
Solution Approach 1:
The patent segments negative samples into multiple groups based on clustering, creating distinct negative prototypes for different groups. This segmentation allows the model to differentiate between various types of non-target sounds more effectively, improving differentiation ability while keeping the overall approach relatively simple.
Solution Approach 2:
The patent applies different treatment to different parts of the data: positive samples form single-class prototypes while negative samples are divided into multiple groups with separate prototypes. This local differentiation strategy enhances the model's ability to distinguish target from non-target sounds without significantly increasing overall complexity.
3Productivity
If single negative prototype is used, then computational efficiency is maintained, but detection of rare events deteriorates
Solution Approach 1:
The patent divides negative samples into multiple groups and creates separate negative prototypes for each group. This segmentation allows rare events to be better represented in specific negative groups, improving detection reliability while maintaining computational efficiency through the use of multiple compact prototypes rather than processing all negative samples individually.
Solution Approach 2:
The patent uses multiple negative prototypes (excessive compared to traditional single prototype) to better cover the diversity of non-target sounds. This partial redundancy improves rare event detection by ensuring that rare events have appropriate negative counterparts for comparison, while the number of prototypes is controlled to maintain computational efficiency.
Data Source
AI summary
A method of training a prototypical network for sound event detection includes receiving samples of an audio signal that include positive samples corresponding to sound events and negative samples that do not correspond to sound events, determining, based on the positive samples, respective positive prototypes of a plurality of classes of sound events, determining, based on the negative samples, respective negative prototypes for respective groups of the negative samples, each of the negative prototypes corresponding to a combination of a plurality of the negative samples, and generating, based on comparisons between a first sample and the respective positive prototypes and each of the negative prototypes, an output signal that indicates whether the first sample belongs to one of the plurality of classes of sound events.