Audio Scene Classification via Synthetic Image Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks face challenges in classifying audio recordings without accompanying video footage due to variations in time, location, and recording equipment, which complicates the classification of audio spectrograms with multiple sources and equipment-induced harmonics.

Innovation Solution

A computing system employing an audio-visual generative adversarial network (AVGAN) with a vector quantized variational autoencoder (VQ-VAE) architecture, including an encoder, decoder, generator, discriminator, and classifier, that generates synthetic image features from audio data to improve classification accuracy across different locations and equipment.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If neural networks classify audio recordings without video footage, then the system can process audio alone, but classification accuracy deteriorates due to variations in time, location, and recording equipment

Engineering Contradiction:
Improveability to process audio without videoVSAvoidclassification accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent creates synthetic audio spectrograms that copy the visual characteristics of real audio scenes. These synthetic spectrograms are generated by converting audio recordings into image-like representations that capture spatial and temporal patterns, allowing the classifier to process audio without requiring actual video footage while maintaining classification accuracy through the synthesized visual features

Inventive Principle:
Principle #26Copying

2Device complexity

If audio spectrograms are processed directly, then the system is simple, but classification performance deteriorates due to multiple sources and equipment-induced harmonics

Engineering Contradiction:
Improvesystem simplicityVSAvoidclassification performance
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent introduces an intermediary transformation layer that converts audio spectrograms into synthetic image features. This intermediary step processes the raw audio data through a generator network that creates enhanced representations, filtering out noise and harmonics while preserving essential scene characteristics, thereby improving classification performance without significantly increasing overall system complexity

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms audio parameters into visual parameter space by generating synthetic images from audio spectrograms. This parameter transformation recasts the classification problem in terms of visual features that are more robust to variations in recording conditions, allowing the system to maintain simplicity while achieving superior classification performance through the transformed feature space

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If the system uses synthetic image features, then classification accuracy improves for unknown locations, but the system complexity increases

Engineering Contradiction:
Improveclassification accuracy for unknown locationsVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates a universal transformation framework that handles both known and unknown locations through the same synthetic image generation process. The generator network learns to create generic visual representations that work across diverse environments, providing multi-functional capability to classify both familiar and unfamiliar scenes without requiring location-specific training data or models, thereby improving accuracy while managing complexity through a single unified approach

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4091167B1Classifying audio scene using synthetic image features
Publication Date: 2024.08.28 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4091167B1 patent drawingFigure 1
  • EP4091167B1 patent drawingFigure 2
  • EP4091167B1 patent drawingFigure 3

AI summary

A computing system includes an encoder that receives an input image and encodes the input image into real image features, a decoder that decodes the real image features into a reconstructed image, a generator that receives first audio data corresponding to the input image and generates first synthetic image features from the first audio data, and receives second audio data and generates second synthetic image features from the second audio data, a discriminator that receives both the real and synthetic image features and determines whether a target feature is real or synthetic, and a classifier that classifies a scene of the second audio data based on the second synthetic image features.