False Invocation Rejection Model for Speech Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech processing systems face issues with false invocation of speech processing skills due to false wakes and misinterpretation of user intents, leading to unintended actions being performed.
Innovation Solution
A false invocation rejection model is deployed downstream of natural language understanding (NLU) and automatic speech recognition (ASR) processing, using semantic rich feature data including confidence scores to classify the likelihood of false invocations, and configurable thresholds based on the importance of the skill action to determine whether to reject or confirm the invocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If speech processing systems activate upon detection of a spoken wakeword, then the system can respond to user commands, but false wakes occur leading to unintended actions
Solution Approach 1:
A false invocation rejection model is introduced as an intermediary component between the wakeword detection and skill execution. This model analyzes semantic features of the utterance and confidence scores to determine whether to reject the invocation, acting as a mediator that filters out false wakes while preserving legitimate user commands
Solution Approach 2:
The system performs preliminary analysis of the utterance by the false invocation rejection model before executing any skill actions. This preliminary evaluation assesses the likelihood of false invocation based on semantic features and confidence scores, preventing unintended actions before they occur
2Measurement precision
If the system performs natural language processing to translate spoken requests, then semantic interpretation is achieved, but misinterpretation of user intents occurs
Solution Approach 1:
The false invocation rejection model provides feedback on the reliability of semantic interpretation by analyzing confidence scores from the natural language processing pipeline. When confidence is low or semantic features suggest misinterpretation, the model rejects the invocation, creating a feedback loop that prevents acting on incorrect intent recognition
Solution Approach 2:
The rejection model serves as an intermediary validation layer between NLU processing and skill execution. It independently evaluates the semantic interpretation using multiple features including confidence scores, utterance characteristics, and contextual information to verify the accuracy of intent recognition before allowing action
3Reliability
If the system rejects all low-confidence invocations, then false invocations are reduced, but legitimate commands may be blocked
Solution Approach 1:
The system dynamically adjusts the rejection threshold based on multiple parameters including confidence scores, utterance characteristics, skill importance, and contextual factors. Rather than using a fixed threshold, the model evaluates the combined weight of multiple parameters to make rejection decisions, allowing flexible adaptation to different situations
Solution Approach 2:
The rejection criteria are customized based on the specific skill and context. Different skills have different importance weights, and the model adjusts its sensitivity accordingly. Critical skills with higher importance require stronger confidence before rejection, while less critical skills may be more readily rejected, creating locally optimized decision-making for different parts of the system
Data Source
AI summary
Devices and techniques are generally described for rejecting false invocations of speech processing skills. In various examples, utterance data comprising automatic speech recognition (ASR) data and natural language understanding (NLU) data may be received. In some examples, ASR confidence data indicating a confidence level of the ASR data may be received. In further examples, NLU confidence data indicating a confidence level of the NLU data may be received. A machine learning model may determine, based at least in part on the ASR confidence data and the NLU confidence data, first false invocation data indicating a likelihood of false invocation of a speech processing skill. In some examples, a first directive may be sent to the speech processing system based at least in part on the first false invocation data. The first directive may be effective to cause the speech processing system to end a current dialog session.