Speech Endpointing via Text Sample Ratio Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional endpointers in natural language processing systems often inaccurately determine the start and end of speech utterances, leading to incomplete processing of voice inputs due to incorrect segmentation based on pause duration, resulting in undesirable or inaccurate outputs.
Innovation Solution
A computing device uses a continuous speech recognizer to incrementally transcribe user audio and compares the transcription to a collection of text samples to determine if it represents a complete or incomplete query, adjusting microphone state accordingly by calculating a ratio of matching text samples with and without additional terms to classify the utterance as complete or incomplete.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional endpointers use pause duration to determine utterance boundaries, then the system can operate with simple logic, but the endpointing accuracy deteriorates leading to incomplete or incorrect phrase segmentation
Solution Approach 1:
The patent introduces an intermediary component (endpointing system) that sits between the speech recognizer and the natural language processing system. This intermediary uses multiple signals (pause detection, word boundary analysis, and confidence scores) to mediate the determination of utterance boundaries, thereby improving endpointing accuracy without requiring complete system redesign.
Solution Approach 2:
The system changes multiple parameters simultaneously to improve endpointing accuracy: it analyzes pause duration thresholds, word boundary characteristics, and confidence score thresholds. By adjusting and combining these parameters, the system achieves more accurate endpointing decisions while maintaining manageable complexity through modular implementation.
2Reliability
If the system waits for complete utterances before processing, then processing accuracy improves, but the response time deteriorates
Solution Approach 1:
The system performs preliminary endpointing assessment during speech recognition using available signals (pauses, word boundaries, confidence scores). This preliminary action allows the system to identify potential utterance boundaries in advance, enabling faster processing decisions without sacrificing accuracy, as the assessment is continuously updated as speech is recognized.
Solution Approach 2:
The endpointing determination is dynamic rather than static. The system continuously evaluates multiple signals and adjusts the endpointing decision as more speech data becomes available. This dynamic approach allows the system to process complete utterances accurately while minimizing waiting time by making decisions as soon as sufficient evidence is gathered.
3Reliability
If the system uses multiple signals for endpointing detection, then endpointing reliability improves, but the computational complexity increases
Solution Approach 1:
The endpointing system is segmented into independent signal analysis components: pause detection module, word boundary analysis module, and confidence score evaluation module. Each component processes one type of signal independently, improving reliability through multiple signals while managing computational complexity through modular, parallel processing that avoids redundant computations.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speech endpointing based on word comparisons are described. In one aspect, a method includes the actions of obtaining a transcription of an utterance. The actions further include determining, as a first value, a quantity of text samples in a collection of text samples that (i) include terms that match the transcription, and (ii) do not include any additional terms. The actions further include determining, as a second value, a quantity of text samples in the collection of text samples that (i) include terms that match the transcription, and (ii) include one or more additional terms. The actions further include classifying the utterance as a likely incomplete utterance or not a likely incomplete utterance based at least on comparing the first value and the second value.