User-Specific Audio Feature Detection for False Trigger Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio feature detection systems in personal assistant devices suffer from high false negative and false positive rates due to variations in user speech accents, dialects, and speech manner, leading to incorrect recognition and frustrating user interactions.
Innovation Solution
Implement a user-specific perturbation generator model trained through adversarial training to generate perturbations that account for individual speech patterns, improving the accuracy of audio feature detection by reducing false negatives and positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional audio feature detection is used without user-specific adaptation, then device complexity is low, but measurement precision deteriorates due to high false negative and false positive rates
Solution Approach 1:
The audio feature detection system is segmented into two distinct components: a generic audio feature detection model for baseline detection and a user-specific perturbation generator model for personalized adaptation. This segmentation allows the system to maintain low complexity for general use while adding specialized functionality only when needed, resolving the contradiction between detection accuracy and device complexity.
Solution Approach 2:
The user-specific perturbation generator model is trained in advance during an enrollment phase using positive audio samples from the user. This preliminary action prepares the model before actual audio feature detection occurs, enabling the system to achieve high measurement precision without increasing real-time detection complexity.
2Measurement precision
If user-specific perturbation models are implemented, then measurement precision improves through reduced false positives, but loss of information increases due to additional processing requirements
Solution Approach 1:
The system applies partial action by using the user-specific perturbation generator model selectively rather than for all audio processing. The model is applied only when user-specific adaptation is beneficial, allowing the system to reduce false positives without incurring excessive processing overhead for every audio sample, thus balancing measurement precision improvement with information loss minimization.
3Reliability
If adversarial training is used to train perturbation generators, then reliability improves through better speech pattern recognition, but device complexity increases due to training complexity
Solution Approach 1:
Adversarial training of the perturbation generator is performed as a preliminary action during the enrollment phase, separate from the actual audio feature detection process. This allows the complex training to occur beforehand without affecting real-time detection complexity, enabling the system to achieve high reliability in speech pattern recognition while maintaining simple operational complexity.
4Measurement precision
If acoustic similarity thresholding is applied, then measurement precision improves by distinguishing similar sounds, but loss of time increases due to additional comparison operations
Solution Approach 1:
The acoustic similarity discrimination process is segmented into two stages: a fast initial filtering stage using the generic model, and a more detailed analysis stage using the user-specific perturbation model only when needed. This segmentation reduces the overall processing time by avoiding exhaustive comparisons for all audio samples while maintaining high measurement precision for distinguishing acoustically similar sounds.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
Training a user-specific perturbation generator for an audio feature detection model includes receiving one or more positive audio samples of a user, each of the one or more positive audio samples including an audio feature; receiving one or more negative audio samples of the user, each of the one or more negative audio samples sharing an acoustic similarity with at least one of the one or more positive audio samples; and adversarially training a user-specific perturbation generator model to generate a user-specific perturbation, the training based on the one or more positive audio samples and the one or more negative audio samples. Perturbing audio samples of the user with the user-specific perturbation can cause an audio feature detection model to recognize the audio feature in audio samples that include the audio feature and/or to refrain from recognizing the audio feature in audio samples that do not include the audio feature.