Utterance End-Point Detection Using Sentence and Word Patterns
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies fail to accurately determine the end point of a user's utterance, impacting speech recognition performance.
Innovation Solution
An electronic device is equipped with a processor that recognizes utterances, determines an End Point Detection (EPD) time based on utterance data, including sentence and word patterns, and terminates recognition when no additional input is made within the EPD time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Voice Activity Detection (VAD) using a voice signal is used to determine whether the user's utterance is completed, then the speech recognition system can detect utterance completion, but the accuracy of determining the end point is insufficient
Solution Approach 1:
The patent changes the parameters used for end point detection from simple voice activity signals to multiple linguistic parameters including sentence patterns, word patterns, and utterance completion indicators. By analyzing these diverse parameters and their combinations, the system achieves more accurate end point detection, resolving the contradiction between detection accuracy and recognition reliability.
2Ease of operation
If the recognition is terminated based on a fixed time threshold, then the system operation is simple, but the speech recognition performance is degraded due to inaccurate end point detection
Solution Approach 1:
The patent transforms the fixed time threshold into a dynamic determination process that adapts to different utterance patterns. The system evaluates multiple parameters including sentence patterns, word patterns, and utterance completion indicators to dynamically adjust the end point detection timing. This dynamic approach maintains operational simplicity while significantly improving speech recognition performance.
3Manufacturing precision
If the EPD time is extended to capture complete utterances, then more complete sentences are recognized, but the response time increases and user experience deteriorates
Solution Approach 1:
The patent applies preliminary action by analyzing sentence patterns, word patterns, and utterance completion indicators during the utterance process to predict the end point in advance. This allows the system to terminate recognition at the optimal moment without waiting for the entire utterance to complete, thereby reducing response time while maintaining high utterance completion accuracy.
4Loss of time
If the recognition terminates too early, then the response time is reduced, but incomplete sentences are recognized leading to errors
Solution Approach 1:
The patent implements feedback mechanisms by continuously monitoring multiple parameters including sentence patterns, word patterns, and utterance completion indicators. The system uses this feedback to dynamically adjust the end point detection timing, ensuring that recognition terminates at the optimal moment - early enough to reduce response time but late enough to capture complete utterances accurately.
Data Source
AI summary
A method of determining an End Point Detection (EPD) time and an electronic device for performing the method are provided. The electronic device includes a processor and a memory electrically connected to the processor and configured to store commands executed by the processor. The processor may be configured to recognize an utterance when the commands are executed, determine an End Point Detection (EPD) time for terminating recognition of the utterance based on utterance data stored in the memory, and terminate the recognition of the utterance when the utterance is not additionally input within the EPD time. The utterance data may include sentence data, which includes a pattern related to a sentence type of the utterance, and word data, which includes a word of the utterance.