Dynamic Ad Detection in Long-Form Audio Using ML Transformers

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current audio content delivery systems face challenges in detecting dynamic advertisements within 'long-form' audio content, such as podcasts and news programs, due to variations in duration and content based on location and listener preferences, which complicates text/audio synchronization and playback resumption.

Innovation Solution

A computing system compares audio files using audio signatures and text transcripts to identify advertisement segments, employing a semantic model with a machine learning transformer and segmentation layer to determine advertisement locations within audio files, allowing for robust detection without a known reference file.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If dynamic advertisements are inserted into audio content based on location and listener preferences, then advertising effectiveness and user personalization are improved, but text/audio synchronization and playback resumption become more difficult

Engineering Contradiction:

Improveadvertising personalizationVSAvoidsynchronization complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system performs preliminary detection and segmentation of advertisement segments within audio files before playback. By pre-identifying ad boundaries and characteristics, the system creates a structured map of ad locations that enables efficient synchronization and resumption without real-time processing complexity

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary detection system that creates a separate layer of advertisement metadata independent of the audio content itself. This intermediary structure stores ad segment information that mediates between the dynamic ad content and the playback system, simplifying synchronization by decoupling ad variations from core playback logic

Inventive Principle:

Principle #24Intermediary (Mediator)

2Adaptability or versatility

If advertisement detection is performed without a known reference file, then adaptability to dynamic advertisements is improved, but detection accuracy and reliability decrease

Engineering Contradiction:

Improvedynamic ad detectionVSAvoiddetection accuracy

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The detection system performs self-service by using the audio content itself to train and refine detection models. The system leverages inherent patterns in the audio data, such as transitions between content and advertisements, to automatically identify ad segments without external reference files, maintaining both adaptability and reliability

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system implements feedback mechanisms where detection results are continuously refined based on detected patterns and characteristics. By analyzing detected ad segments and using this information to improve future detections, the system achieves high reliability without requiring pre-existing reference files

Inventive Principle:

Principle #23Feedback

3Measurement precision

If deep learning models are used for advertisement segmentation, then detection precision is improved, but computational resources and processing time increase

Engineering Contradiction:

Improvead segment detection precisionVSAvoidcomputational resource consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies segmentation by dividing the audio file into potential advertisement segments and non-advertisement segments. This segmentation approach allows the deep learning model to process smaller, focused portions of audio data rather than entire files, reducing computational load while maintaining detection precision through targeted analysis of segment boundaries and characteristics

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS12190871B1Deep learning-based automatic detection and labeling of dynamic advertisements in long-form audio content

Publication Date: 2025.01.07 AMAZON TECH INC

AI summary

Techniques and methods are disclosed for detecting long-form audio content in one or more audio files. A computing system receives first audio data corresponding to a first version of an audio file and second audio data corresponding to a second version of the audio file. The computing system generates a first transcript of the first audio data and a second transcript of the second audio data. The computing system compares the first audio data and the second audio data and the first transcript and the second transcript to identify advertisement portions and content portions of the audio data. Using a semantic model based on a machine learning (ML) transformer, the computing system can determine advertisement segments within the advertisement portions, the advertisement segments corresponding to separate advertisements. Information corresponding to the duration and location of the advertisement segments is stored in a data store of the computing system.