Wake-Word Recognition Model Segmentation for Low-Power Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing dialog systems face challenges in accurately recognizing wake-words on low-power devices due to limited compute, memory, and power resources, leading to potential false positives and negatives, which waste energy and impact privacy or fail to process user inputs.
Innovation Solution
A method that shares a wake-word recognition model across a listening device and a processing device, utilizing ensemble classification to combine the wake-word recognition model with a filtering model, improving accuracy while keeping the wake-word recognition model small and efficient.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a wake-word recognition model is implemented on a low-power device, then the device can recognize wake-words locally, but the limited compute, memory, and power resources lead to reduced recognition accuracy and increased false positives/negatives
Solution Approach 1:
The system divides wake-word recognition into two segments: a lightweight model on the low-power listening device for initial detection, and a more sophisticated model on a remote processing device for final verification. This segmentation allows each device to operate within its resource constraints while achieving overall high accuracy through collaborative processing.
Solution Approach 2:
The listening device acts as an intermediary between the user and the processing device. It performs preliminary wake-word detection using a compact model, then forwards candidate detections to the processing device for verification. This intermediary role enables the system to maintain low power consumption at the edge while leveraging cloud-based computational power for accurate final recognition.
2Measurement precision
If the wake-word recognition model is made larger and more sophisticated to improve accuracy, then false positives and negatives are reduced, but the device requires more compute, memory, and power resources
Solution Approach 1:
The recognition model is segmented into two components: a small, efficient model deployed on the listening device for initial wake-word candidate detection, and a larger, more accurate model on the processing device for final verification. This segmentation allows the system to achieve high precision without requiring the entire sophisticated model to reside on the resource-constrained listening device.
Solution Approach 2:
The listening device performs partial wake-word recognition using a simplified model, then transfers only the necessary audio data or features to the processing device for complete verification. This partial action approach allows the system to benefit from both the low-latency local processing and the high accuracy of the full model without requiring the full model on the edge device.
3Reliability
If false positives are increased to ensure no wake-words are missed, then sensitivity improves, but more non-wake-word inputs are incorrectly triggered, wasting power and potentially compromising privacy
Solution Approach 1:
The listening device serves as an intermediary that performs initial sensitivity-based detection, then acts as a filter by forwarding only promising candidates to the processing device for verification. This two-stage approach with the listening device as intermediary allows the system to maintain high sensitivity without suffering from the full impact of false positives, as the processing device verifies candidates before final triggering.
Solution Approach 2:
The listening device performs preliminary wake-word candidate detection using a simple model with high sensitivity settings, then pre-filters the results before sending to the processing device. This preliminary action allows the system to cast a wide net for potential wake-words without committing resources to process every false positive, as the processing device performs final verification only on promising candidates.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Techniques for recognizing a wake-word in a dialog system through the use of (a) a wake-word recognition model to identify wake-word candidates and (b) ensemble classification utilizing the wake-word recognition model along with a filtering model to recognize the wake-word in wake-word candidates. If the wake-word is recognized, the dialog system listens for speech input to process.