Trigger-Word Detection Training Through Self-Labeled Real-World Audio

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training machine-learning models for trigger-word detection is tedious and time-consuming, and existing methods do not effectively leverage real-world data for continuous improvement.

Innovation Solution

A method for training machine-learning models using a training data provider that automatically labels and provides sound-based data for trigger-word detection, utilizing federated learning to aggregate local models and improve recognition of trigger words over time, even in varied utterances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual training data labeling is used for ML model training, then training can be performed, but the process becomes tedious and time-consuming

Engineering Contradiction:
Improvetrigger-word detection accuracyVSAvoidtraining data preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system automatically labels training data by capturing sound-based data from the monitored environment, detecting trigger words using an existing ML model, and using these detections to automatically create labeled training examples. This self-service approach eliminates manual labeling while continuously improving the model through real-world data collection and automatic labeling.

Inventive Principle:
Principle #25Self-service

2Reliability

If existing ML models are used for trigger-word detection, then detection can be performed, but the model cannot effectively leverage real-world data for continuous improvement

Engineering Contradiction:
Improvetrigger-word detection reliabilityVSAvoidmodel adaptability to varied utterances
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system implements a feedback loop where the ML model continuously detects trigger words in real-world audio data, these detections are automatically labeled and used to retrain the model, improving its performance over time. The system adapts to varied utterances by collecting diverse real-world examples and incorporating them into continuous training cycles.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs preliminary actions by automatically collecting and labeling training data from the monitored environment before model retraining. Sound-based data is captured and pre-labeled with trigger word detections, preparing ready-to-use training examples that enable subsequent model improvement without manual intervention.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If manual collection of training data is performed, then training data can be obtained, but the process is tedious and time-consuming

Engineering Contradiction:
Improveamount of training dataVSAvoiddata collection efficiency
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The system automatically collects sound-based data from the monitored environment and self-labels it by detecting trigger words using the ML model. This eliminates the need for manual data collection and labeling, continuously accumulating training data in the background without human intervention while maintaining high productivity.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4248438B1Enabling training of a machine-learning model for trigger-word detection
Publication Date: 2025.07.02 ASSA ABLOY AB
  • EP4248438B1 patent drawingFigure 1~2C
  • EP4248438B1 patent drawingFigure 3A~3B
  • EP4248438B1 patent drawingFigure 4~5

AI summary

It is provided a method for enabling training a machine-learning, ML, model for trigger-word detection, the method being performed in a training data provider (1). The method comprises: receiving (40) sound-based data, the sound-based data being based on sounds captured in a space to be monitored; determining (42) that the sound-based data corresponds to a trigger word, and labelling this sound-based data to correspond to the trigger word; and providing (44) the labelled sound-based data to train the ML model.