Logarithmic Scale Convolutional Neural Network for Audio Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Audio processing systems face reduced accuracy and performance due to information loss during the mapping of frequency domain features from a linear scale to a logarithmic scale, and hardcoded fixed-scale mappings are not optimal for different classification tasks.

Innovation Solution

A logarithmic scale convolutional neural network system learns uniform logarithmic scale convolutions across the frequency spectrum, and a cascaded convolutional neural network system applies variable-sized convolutions, allowing for improved feature extraction and selection of optimal scales for different frequency spectrum parts.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If frequency domain features are mapped from linear scale to logarithmic scale using hardcoded fixed-scale mapping, then the audio processing system can process frequency features, but information is lost during the mapping process and accuracy is reduced

Engineering Contradiction:
Improveclassification accuracyVSAvoidfrequency feature information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent applies dynamics by transitioning from hardcoded fixed-scale mapping to learned dynamic scale mappings. The neural network learns optimal scale transformations adaptively during training, allowing the mapping parameters to adjust dynamically based on the input data characteristics rather than using static predetermined mappings.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent implements parameter changes by making the scale mapping parameters learnable instead than fixed. The system changes from using predetermined mapping parameters to using parameters that are optimized during training, allowing the mapping characteristics to evolve based on the specific task and data distribution.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If uniform logarithmic scale convolutions are applied across the whole frequency spectrum, then the system achieves consistent feature extraction, but different classification tasks and input signal characteristics require different optimal mappings

Engineering Contradiction:
Improvetask-specific optimizationVSAvoidconvolutional mapping complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the frequency spectrum into multiple bands and applying different scale convolutions to different segments. This allows each frequency band to be processed with task-optimized mappings while maintaining overall system coherence, resolving the conflict between uniformity and adaptability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements local quality by allowing different scale convolution parameters to be applied to different frequency bands and tasks. Each local region of the frequency spectrum can have its own optimized mapping characteristics, enabling task-specific adaptation without requiring complete system redesign.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If variable sized convolutions are used to produce different resolutions, then optimal scales can be selected for different parts of frequency spectrum, but computational resources and memory requirements increase

Engineering Contradiction:
Improvefeature extraction accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies dynamics by implementing a cascade architecture where convolutional windows dynamically adjust their sizes based on the frequency band being processed. Lower frequency bands use larger windows while higher frequency bands use smaller windows, optimizing computational resources for each region's specific requirements.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent implements partial action by applying variable sized convolutions only where necessary in the frequency spectrum rather than uniformly across all bands. This selective approach maintains high accuracy where needed while reducing overall computational burden.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3440598B1Frequency based audio analysis using neural networks
Publication Date: 2023.11.08 GOOGLE LLC
  • EP3440598B1 patent drawingFigure 1
  • EP3440598B1 patent drawingFigure 2
  • EP3440598B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for frequency based audio analysis using neural networks. One of the methods includes training a neural network that includes a plurality of neural network layers on training data, wherein the neural network is configured to receive frequency domain features of an audio sample and to process the frequency domain features to generate a neural network output for the audio sample, wherein the neural network comprises (i) a convolutional layer that is configured to map frequency domain features to logarithmic scaled frequency domain features, wherein the convolutional layer comprises one or more convolutional layer filters, and (ii) one or more other neural network layers having respective layer parameters that are configured to process the logarithmic scaled frequency domain features to generate the neural network output.