End of Utterance Detection via Score Stability Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face errors in detecting the end of utterance, particularly in silent periods before speech, delays between speech segments, and noise-induced failures, which affect recognition accuracy and usability in mobile devices.
Innovation Solution
A method that stabilizes recognition results by processing best state scores and token scores, using pre-calculated values to determine end of utterance by comparing sums and slopes to threshold values, reducing errors and improving detection reliability in noisy environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If stability check of recognition result is used for EOU detection, then decoding can be stopped when recognition result stabilizes, but errors occur in silent periods before speech, during speech delays, and in noisy environments
Solution Approach 1:
The patent segments the EOU detection process into multiple independent components: stability check of recognition result, best state score evaluation, and best token score evaluation. Each component operates independently and contributes to the final EOU decision, allowing the system to leverage multiple indicators rather than relying on a single metric that may fail in certain conditions.
Solution Approach 2:
The patent implements feedback mechanisms where the speech recognizer continuously monitors its own recognition results, best state scores, and best token scores during decoding. This self-feedback allows the system to detect when the recognition result has stabilized and when EOU should be detected, improving both efficiency and reliability through continuous self-evaluation.
2Device complexity
If pre-calculated state and token scores are used for EOU detection, then computational complexity is reduced for mobile devices, but detection reliability in noisy environments must be maintained
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing state scores and token scores during the speech recognition setup phase. These pre-calculated values are then reused during EOU detection without requiring complex real-time computations, significantly reducing the computational burden on mobile devices while maintaining detection accuracy through the use of these pre-computed metrics.
Data Source
AI summary
The present invention relates to speech recognition systems, especially to arranging detection of end-of utterance in such systems. A speech recognizer of the system is configured to determine whether recognition result determined from received speech data is stabilized. The speech recognizer is configured to process values of best state scores and best token scores associated with frames of received speech data for end of utterance detection purposes. Further, the speech recognizer is configured to determine whether end of utterance is detected or not, based on the processing, if the recognition result is stabilized.