Hierarchical Keyword Detection for False Alarm Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing keyword detection systems in voice-based interactions suffer from high false alarm rates, leading to unnecessary device activation, power consumption, and privacy concerns, especially when processing out-of-domain audio, without effectively addressing these issues with small model sizes and computational efficiency.
Innovation Solution
A successive refinement-based keyword detection system with multiple layers that classify audio inputs hierarchically, including a backbone model for common features and specialized classifiers for speech, keyword-like speech, and specific keywords, reducing false alarms without increasing model size or computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a keyword detection system processes audio continuously to maintain high detection rate, then detection capability is improved, but false alarm rate increases
Solution Approach 1:
The keyword detection system is segmented into multiple independent layers: a first layer detects speech presence, a second layer detects keyword-like speech, and a third layer performs specific keyword classification. This segmentation allows each layer to specialize in its specific task, improving overall detection accuracy while reducing false alarms through progressive filtering
Solution Approach 2:
The system performs preliminary detection actions in the first layer by identifying speech presence before proceeding to keyword-specific detection. This preliminary action filters out non-speech audio early in the processing pipeline, preventing false alarms from non-speech inputs while maintaining high detection rates for actual speech containing keywords
2Reliability
If a keyword detection system uses larger models and datasets to reduce false alarms, then false alarm rate is reduced, but device power consumption increases
Solution Approach 1:
The detection system is divided into multiple layers with progressively specialized functions. The first layer handles basic speech detection, the second layer handles keyword-like speech detection, and the third layer handles specific keyword classification. This segmentation allows computationally intensive operations to be performed only when necessary, reducing overall power consumption while maintaining low false alarm rates
Solution Approach 2:
The system performs only the necessary detection actions at each layer based on the input characteristics. The first layer performs basic speech detection, and only if speech is detected does the system proceed to the more computationally intensive second and third layers. This partial action approach avoids unnecessary processing of non-speech audio, reducing power consumption while maintaining detection accuracy
3Measurement precision
If a keyword detection system uses larger models to improve detection accuracy, then detection accuracy is improved, but model size increases
Solution Approach 1:
The keyword detection model is segmented into multiple layers with specialized functions: the first layer detects speech presence, the second layer detects keyword-like speech, and the third layer performs specific keyword classification. Each layer uses a model size appropriate for its specific task, avoiding the need for a single large model to handle all detection functions, thus reducing overall model complexity while maintaining high detection accuracy
Data Source
AI summary
A method includes extracting, using a keyword detection model, audio features from audio data. The method also includes processing the audio features by a first layer of the keyword detection model configured to predict a first likelihood that the audio data includes speech. The method also includes processing the audio features by a second layer of the keyword detection model configured to predict a second likelihood that the audio data includes keyword-like speech. The method also includes processing the audio features by a third layer of the keyword detection model configured to predict a third likelihood, for each of a plurality of possible keywords, that the audio data includes the keyword. The method also includes identifying a keyword included in the audio data. The method also includes generating instructions to perform an action based at least in part on the identified keyword.