Wake-Word Detection with Post-ASR False Wake-Up Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
False wake-ups in voice assistants lead to bad user experiences, resource wastage, and privacy concerns, especially with seamless registration processes that lack audio training samples, resulting in decreased accuracy of wake-word detection systems.
Innovation Solution
A system and method for post-ASR false wake-up suppression using multiple machine learning models to analyze speech signals, including an audio wake-up classifier, ASR model, and false wake-up suppression classifier, leveraging audio and context features to delay and improve wake-up decision-making.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If seamless registration is used to eliminate audio training sample collection, then user experience is improved by eliminating a registration step, but wake-word detection accuracy deteriorates due to lack of training data
Solution Approach 1:
The wake-word detection system is segmented into multiple independent models: a first machine learning model for initial wake-word detection, a second machine learning model for false wake-up suppression, and an automatic speech recognition model. Each model operates independently with its own specialized function, allowing the system to maintain high accuracy without requiring audio training samples during registration.
Solution Approach 2:
Automatic speech recognition is introduced as an intermediary component between audio input and wake-word detection decisions. The ASR model processes the audio signal and provides contextual information to the second machine learning model, enabling more accurate false wake-up suppression without requiring user-provided training data.
2Productivity
If wake-word detection sensitivity is increased to catch all potential wake-words, then wake-up rate is improved, but false wake-up rate increases leading to resource wastage and privacy concerns
Solution Approach 1:
The first machine learning model performs preliminary wake-word detection to identify potential wake-up events. This preliminary action filters out obviously non-wake-up audio signals before they reach the second model, allowing the system to maintain high sensitivity while reducing false positives through subsequent verification.
Solution Approach 2:
The system implements a feedback mechanism where the second machine learning model receives inputs from both the first model and the automatic speech recognition system. The ASR output provides feedback about the actual speech content, enabling the second model to adjust its decision and suppress false wake-ups that the first model may have incorrectly identified.
Data Source
AI summary
A method includes obtaining a speech signal. The method also includes predicting a first likelihood of a wake word or phrase being spoken in the speech signal using a first machine learning model trained to receive the speech signal as input. The method further includes, responsive to the first likelihood exceeding a first threshold, performing automatic speech recognition on the speech signal to determine a textual representation of the speech signal. The method also includes predicting a second likelihood of the wake word or phrase being spoken in the speech signal using a second machine learning model trained to receive at least one of the textual representation, audio features associated with the speech signal, and context features associated with the electronic device. In addition, the method includes, responsive to the second likelihood exceeding a second threshold, generating instructions to perform an action requested in the speech signal.