Context-Sensitive Wake Word Detection Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Digital assistants often mistakenly activate when users casually mention their wake words during conversations, leading to unnecessary wake-ups and user annoyance, as they fail to differentiate between vocative and non-vocative uses of the wake word.
Innovation Solution
A machine-trained neural network, such as a recurrent neural network (RNN) or long short-term memory (LSTM) network, is trained to distinguish between vocative and non-vocative uses of the wake expression by utilizing both grammatical and prosodic components, allowing it to contextually identify when the wake word is being used to invoke the assistant.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If simple speech recognition processes are used to detect wake words, then the digital assistant can respond to wake word invocations, but it also activates unnecessarily when the wake word is mentioned casually in conversation
Solution Approach 1:
The speech recognition process is segmented into two distinct stages: a first simple speech recognition process that runs continuously with low computational requirements, and a second more complex speech recognition process that is triggered only when the first process detects a potential wake word. This segmentation allows the system to maintain high wake word detection accuracy while avoiding the continuous computational burden of the more complex process, thereby reducing false activations without significantly increasing overall system complexity.
2Reliability
If the digital assistant activates on every wake word-like sequence, then it ensures no invocations are missed, but it causes user annoyance due to false activations
Solution Approach 1:
A prosodic analysis component is introduced as an intermediary between the first speech recognition process and the digital assistant's activation. This intermediary analyzes the prosodic features (intonation, stress, rhythm) of the detected wake word candidate and determines whether it represents a genuine invocation or a casual mention. By adding this intermediary layer, the system maintains high reliability in detecting actual invocations while significantly reducing false activations that cause user annoyance.
3Reliability
If the digital assistant uses context-sensitive detection, then it reduces false activations, but it increases processing complexity and computational requirements
Solution Approach 1:
The system performs preliminary wake word detection using a simple, low-complexity speech recognition process before engaging the more complex context-sensitive analysis. This preliminary action filters out obvious non-invocations early in the process, allowing the system to apply context-sensitive detection only to ambiguous cases. As a result, the system achieves high wake word invocation accuracy while minimizing the overall computational burden and processing complexity.
Data Source
AI summary
Some embodiments provide a method of training a MT network to detect a wake expression that directs a digital assistant to perform an operation based on a request that follows the expression. The MT network includes processing nodes with configurable parameters. The method iteratively selects different sets of input values with known sets of output values. Each of a first group of input value sets includes a vocative use of the expression. Each of a second group of input value sets includes a non-vocative use of the expression. For each set of input values, the method uses the MT network to process the input set to produce an output value set and computes an error value that expresses an error between the produced output value set and the known output value set. Based on the error values, the method adjusts configurable parameters of the processing nodes of the MT network.