Cascade Hotword Detection with Personalized Negative Hotwords
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech-enabled devices face challenges in accurately distinguishing between intended hotwords and false positives due to variations in user speech, accents, and acoustic environments, leading to inefficient and user-unfriendly wake-up processes.
Innovation Solution
A cascade hotword detection architecture with a first stage detector on the device and a second stage detector on a remote system, combined with a negative hotword classifier, personalizes hotword recognition by identifying and updating the first stage detector to avoid false positives based on user-specific and device-specific negative hotwords.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a first stage hotword detector is used to detect hotwords in streaming audio, then hotword detection speed is improved, but false positive rate increases
Solution Approach 1:
The hotword detection system is segmented into two stages: a first stage detector that operates quickly with lower accuracy to maintain speed, and a second stage detector that operates more slowly with higher accuracy to reduce false positives. This segmentation allows the system to achieve both fast response and high reliability by distributing the detection task across multiple specialized components.
Solution Approach 2:
A negative hotword classifier acts as an intermediary component between the first stage detector and the second stage detector. It processes audio segments that triggered false positives in the first stage, analyzing them to identify and suppress negative hotwords before they reach the second stage detector, thereby reducing false positives while maintaining the speed benefits of the first stage approach.
2Reliability
If a second stage hotword detector is used to confirm hotword detection, then false positive rate is reduced, but processing time increases
Solution Approach 1:
The negative hotword classifier performs preliminary action by pre-processing audio segments that have triggered false positives in the first stage detector. It identifies and suppresses negative hotwords before the audio segments reach the second stage detector, thereby reducing the workload and processing time of the second stage while maintaining its ability to reduce false positives.
Solution Approach 2:
Instead of applying the resource-intensive second stage detector to all audio segments, the system applies it only partially to segments that have already been flagged by the first stage detector. The negative hotword classifier further refines this by filtering out segments containing negative hotwords, ensuring the second stage processes only the most relevant cases, thus minimizing processing time while maintaining high reliability.
3Measurement precision
If the hotword detector is updated with personalized negatives, then detection accuracy is improved, but device complexity increases
Solution Approach 1:
The system implements self-service by automatically collecting false positive audio segments, classifying them as negative hotwords, and updating the first stage detector without requiring manual intervention. The negative hotword classifier continuously learns from new false positives and automatically integrates this knowledge into the detection system, improving accuracy while keeping the complexity management automated rather than manual.
Solution Approach 2:
A feedback loop is established where false positives detected by the first stage are fed into the negative hotword classifier, which then updates the detector with learned negative hotword patterns. This continuous feedback mechanism allows the system to automatically improve its detection accuracy over time by learning from its mistakes, managing complexity through automated adaptation rather than manual configuration.
Data Source
AI summary
A method for adapting hotword recognition includes receiving audio data characterizing a hotword event detected by a first stage hotword detector in streaming audio captured by a user device. The method also includes processing, using a second stage hotword detector, the audio data to determine whether a hotword is detected by the second stage hotword detector in a first segment of the audio data. When the hotword is not detected by the second stage hotword detector, the method includes, classifying the first segment of the audio data as containing a negative hotword that caused a false detection of the hotword event in the streaming audio by the first stage hotword detector. Based on the first segment of the audio data classified as containing the negative hotword, the method includes updating the first stage hotword detector to prevent triggering the hotword event in subsequent audio data that contains the negative hotword.