Unsupervised Audio Embedding Triplet Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The high cost and limitations of manually generating comprehensive labels for training artificial neural networks to classify audio recordings, particularly for non-speech sounds, which are often not adequately represented in exclusive label sets.

Innovation Solution

An unsupervised method is employed to generate mappings between audio segments and feature vectors in a semantically-encoded n-dimensional feature space, using triplet sampling and loss functions to train the neural networks, allowing for the identification of similar sounds and bootstrapping labels from a small set of supervised examples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If manual labels are used to train neural networks for audio classification, then the classifier can be trained with supervised learning, but the cost increases and the label set becomes limited to an exclusive set that may not cover all sound contents

Engineering Contradiction:
Improveclassifier training qualityVSAvoidmanual labeling cost and complexity
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The system uses unsupervised learning to automatically generate feature representations from audio data without requiring manual labels. The neural network learns to cluster and represent audio segments based on their inherent characteristics, enabling the system to serve itself by discovering patterns without human intervention for labeling.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent creates synthetic training data by generating pseudo-labels from unsupervised clustering results. These synthetic labels copy the structure of real labeled data, allowing the system to train classifiers using automatically generated examples that mirror the format and utility of manually created labels.

Inventive Principle:
Principle #26Copying

2Ease of operation

If an exclusive set of labels is used for manual labeling, then the labeling process is simplified with predefined categories, but the representation of all possible sound contents becomes inadequate

Engineering Contradiction:
Improvelabeling process simplicityVSAvoidcoverage of sound contents
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system dynamically generates labels based on the actual audio data patterns rather than using a static predefined set. The unsupervised learning process adapts the label space to the specific characteristics of the audio content, allowing new categories to emerge naturally from the data without requiring predetermined classifications.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent transitions from a flat, fixed-label classification system to a multi-dimensional feature space where audio segments are represented by their inherent characteristics. This dimensional expansion allows for more nuanced and comprehensive representation of sound contents beyond the limitations of exclusive label sets.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Productivity

If unsupervised learning is used to generate feature mappings, then manual labeling is reduced and comprehensive representation is achieved, but the complexity of training and feature space mapping increases

Engineering Contradiction:
Improvelabeling efficiencyVSAvoidtraining process complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system segments audio recordings into smaller audio segments and processes them independently through the neural network. This segmentation allows the complex unsupervised learning task to be broken down into manageable units, where each segment is processed and represented individually, then aggregated to form comprehensive feature mappings for the entire audio content.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3688750B1Unsupervised learning of semantic audio representations
Publication Date: 2024.03.13 GOOGLE LLC
  • EP3688750B1 patent drawingFigure 1A~1B
  • EP3688750B1 patent drawingFigure 2
  • EP3688750B1 patent drawingFigure 3A~3C

AI summary

Methods are provided for generating training triplets that can he used to train multidimensional embeddings to represent the semantic content of non-speech sounds present in a corpus of audio recordings. These training triplets can be used with a triplet loss function to train the multidimensional embeddings such that the embeddings can be used to cluster the contents of a corpus of audio recordings, to facilitate a query-by-example lookup from the corpus, to allow a small number of manually-labeled audio recordings to be generalized, or to facilitate some other audio classification task. The triplet sampling methods may be used individually or collectively, and each represent a respective heuristic about the semantic structure of audio recordings.