Wake-up Model Training Using False Activation Counterexamples

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing device wake-up methods frequently result in false awakenings, which are costly and affect travel safety, as they require additional corpus recording and are not effectively suppressed.

Innovation Solution

A method where voice information during false awakenings is used as counterexample samples to train a wake-up model, eliminating the need for additional corpus recording and automatically suppressing false awakenings, thereby improving wake-up accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional wake-up models are used, then device can be woken up by voice commands, but false wake-up events occur frequently

Engineering Contradiction:
Improvewake-up accuracyVSAvoidfalse wake-up events
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent converts false wake-up events from harmful errors into beneficial training data. By collecting voice recordings during false wake-up events and using them as negative samples in model training, the system transforms mistakes into learning opportunities, continuously improving wake-up accuracy while reducing future false activations.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The patent implements a feedback loop where wake-up events (including false wake-ups) are collected, analyzed, and used to retrain the wake-up model. This closed-loop system continuously improves model performance by feeding back real-world performance data, particularly using false wake-up cases to adjust decision boundaries and reduce future errors.

Inventive Principle:
Principle #23Feedback

2Reliability

If additional corpus recording is performed to improve wake-up accuracy, then model performance improves, but development cost and time increase

Engineering Contradiction:
Improvewake-up accuracyVSAvoidcorpus recording requirements
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent enables the system to self-improve without external intervention. The wake-up model automatically collects its own performance data during operation, including false wake-up events, and uses this data for self-training and retraining. This eliminates the need for manual corpus recording and external data collection efforts.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent changes the training data parameters by incorporating false wake-up voice recordings as negative samples. Instead of requiring additional carefully recorded corpus data, the system utilizes previously wasted or discarded false wake-up events, transforming them into valuable training material that adjusts model parameters and improves accuracy.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3923272B1Method and apparatus for adapting a wake-up model
Publication Date: 2023.05.24 APOLLO INTELLIGENT CONNECTIVITY (BEIJING) TECH CO LTD
  • EP3923272B1 patent drawingFigure 1
  • EP3923272B1 patent drawingFigure 2
  • EP3923272B1 patent drawingFigure 3

AI summary

Embodiments of the present disclosure disclose a method and apparatus for outputting information, which can be used in an intelligent traffic scene. An embodiment of the method includes acquiring voice information received within a preset time period before a device is awakened, where the device is provided with a wake-up model for outputting preset response information when a preset wake-up word is received; performing speech recognition on the voice information to obtain a recognition result; extracting feature information of the voice information in response to determining that the recognition result does not include a preset wake-up word; generating a counterexample training sample according to the feature information; and training the wake-up model using a counter-example training sample, and outputting the trained wake-up model. In this embodiment, the voice information during false wakeup can be used as an example to train a wake-up model, no additional corpus recording is required, cost is saved, false wake-up can be automatically suppressed, and accuracy of wakeup is improved.