Wakeword Detection Model Update via Error Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face issues with false-positive and false-negative detections of wakewords, leading to diminished usability and user frustration, as they fail to accurately differentiate between intended wakewords and background noise or misinterpretations.
Innovation Solution
The system updates a trained model based on detected false positives and negatives by back-propagating differences between expected and actual wakeword representations in audio data, using techniques like Hidden Markov Models and deep neural networks, and aggregates this information across devices to improve wakeword detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition systems use fixed wakeword detection models, then device complexity is reduced and ease of operation is improved, but detection accuracy deteriorates leading to false positives and false negatives
Solution Approach 1:
The speech recognition system automatically updates its own wakeword detection model by collecting false positive and false negative examples during operation, training the model with these examples, and deploying the updated model without requiring external intervention. This self-service approach improves detection accuracy while minimizing the complexity burden on users.
Solution Approach 2:
The system implements a feedback loop where detection errors (false positives and false negatives) are collected and used to retrain the wakeword detection model. This continuous feedback mechanism allows the system to learn from its mistakes and improve accuracy over time, resolving the contradiction between maintaining simple operation and achieving high detection precision.
2Measurement precision
If the system continuously updates the wakeword detection model with new data, then detection accuracy improves, but loss of time increases due to training and deployment cycles
Solution Approach 1:
The system collects and stores false positive and false negative examples during normal operation, preparing training data in advance. When model updates are needed, the pre-collected data can be immediately used for training, reducing the time loss associated with data collection and preparation during the update process.
Solution Approach 2:
The system selectively updates the model only when sufficient training examples are collected or when performance degradation is detected, rather than continuously training. This partial action approach balances accuracy improvement with time efficiency, avoiding unnecessary training cycles while still achieving meaningful accuracy gains.
3Measurement precision
If the system collects and processes more audio data for model training, then wakeword detection accuracy improves, but loss of information increases due to potential privacy concerns and data security requirements
Solution Approach 1:
The system processes and trains on audio data locally on the device rather than transmitting raw audio data to remote servers. This local processing approach maintains detection accuracy by using sufficient training data while protecting user privacy and security by keeping sensitive information on-device, effectively resolving the contradiction between accuracy improvement and information loss.
Solution Approach 2:
The system extracts only the essential features and patterns from audio data needed for model training, rather than storing or processing complete raw audio recordings. This extraction approach enables accurate model training while minimizing the retention of sensitive personal information, thereby reducing privacy and security concerns.
Data Source
AI summary
A device monitors audio data for a predetermined and/or user-defined wakeword. The device detects an error in detecting the wakeword in the audio data, such as a false-positive detection of the wakeword or a false-negative detection of the wakeword. Upon detecting the error, the device updates a model trained to detect the wakeword to create an updated trained model; the updated trained model reduces or eliminates further errors in detecting the wakeword. Data corresponding to the updated trained model may be collected by a server from a plurality of devices and used to create an updated trained model aggregating the data; this updated trained model may be sent to some or all of the devices.