Relaxed Instance Frequency Normalization for Neural Audio Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for processing audio content using artificial neural networks face challenges in effectively extracting usable information, particularly due to the concentration of statistical information, which increases the difficulty of classification tasks.
Innovation Solution
The method involves training a neural network using a relaxed feature-normalized data set, where audio samples are normalized in the frequency domain with additional normalization to relax the loss of discriminative information, allowing for domain generalization and improved classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional frequency normalization is applied to audio data, then domain generalization is improved, but discriminative information is lost
Solution Approach 1:
The patent applies relaxed instance frequency normalization that modifies the normalization parameters by introducing a relaxation factor α (where 0 < α ≤ 1). This parameter change allows the normalization to be less aggressive, preserving more discriminative information while still achieving domain generalization. The normalized feature is computed as: x_normalized = (x - μ_f) / (σ_f + ε), where the statistics μ_f and σ_f are calculated with relaxation applied across frequency bins.
Solution Approach 2:
Instead of applying full instance normalization across all frequency dimensions, the patent applies partial normalization by selectively relaxing the normalization strength. The relaxation factor α controls the degree of normalization applied, allowing the system to perform partial normalization that suffices for domain generalization without excessively removing discriminative information. This partial action approach balances the competing requirements.
2Measurement precision
If multiple classification models are used to handle different domains, then classification accuracy is improved, but computational resources and power consumption increase
Solution Approach 1:
The patent creates a single universal neural network model that can handle multiple audio domains through relaxed instance frequency normalization. This normalized model serves multiple functions: it performs domain generalization, maintains classification accuracy across different domains, and eliminates the need for separate domain-specific models. The single model approach reduces computational overhead and power consumption while maintaining the accuracy benefits of multiple specialized models.
Data Source
AI summary
Techniques and apparatus for training a neural network to classify audio into one of a plurality of categories and using such a trained neural network. An example method generally includes receiving a data set including a plurality of audio samples. A relaxed feature-normalized data set is generated by normalizing each audio sample of the plurality of audio samples. A neural network is trained to classify audio into one of a plurality of categories based on the relaxed feature-normalized data set, and the trained neural network is deployed.