Sound Reception Termination Detection Using Semantic Relevance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing dialogue robots inaccurately determine the termination of human speech, leading to errors and poor user experience, especially in continuous sound scenarios, as they rely solely on voice activity detection and fail to recognize follow-up information.
Innovation Solution
A method and apparatus that perform voice activity detection and semantic relevance detection using deep learning on current and next sound clips, followed by a weighted calculation and comparison with a preset threshold to determine sound reception termination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only voice activity detection is used to judge speech termination, then the device complexity is low, but the measurement precision of speech termination is poor
Solution Approach 1:
The patent merges voice activity detection (VAD) with semantic relevance detection to form a hybrid detection system. The terminal device simultaneously performs VAD to identify speech boundaries and semantic relevance detection to determine whether subsequent sound clips contain follow-up information, combining both detection results to judge speech termination accurately. This resolves the contradiction by integrating multiple detection mechanisms while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent introduces a semantic relevance detection module as an intermediary between voice activity detection and speech termination judgment. This intermediary analyzes the semantic content of sound clips to determine whether they contain follow-up information, acting as a bridge that enhances detection accuracy without directly increasing the complexity of the core VAD system. The semantic relevance detector processes audio data and provides additional context to the termination judgment logic.
2Reliability
If voice activity detection alone is used, then the processing speed is fast, but the reliability of speech termination judgment is poor
Solution Approach 1:
The patent applies preliminary action by performing semantic relevance detection on sound clips in advance to identify potential follow-up information before final speech termination judgment is made. The system pre-processes audio data to detect semantic relationships, so when termination judgment is needed, the reliability is enhanced because the semantic context has already been analyzed. This reduces the need for re-processing and improves overall judgment reliability without proportionally increasing processing time.
Solution Approach 2:
The patent implements feedback mechanisms where the results of semantic relevance detection are fed back into the speech termination judgment process. The system continuously monitors both VAD outputs and semantic relevance results, using this feedback to refine termination decisions. This feedback loop ensures that even if VAD alone would indicate termination, the system can detect semantic continuity and adjust the judgment accordingly, significantly improving reliability while maintaining efficient processing through iterative refinement.
3Measurement precision
If semantic relevance detection by deep learning is added, then the measurement precision improves, but the use of energy increases
Solution Approach 1:
The patent applies partial action by selectively applying semantic relevance detection only when VAD indicates potential speech termination or when contextual cues suggest follow-up information may be present. Rather than continuously performing computationally intensive deep learning semantic analysis on all audio data, the system triggers semantic detection only in relevant scenarios, significantly reducing energy consumption while maintaining high detection accuracy when it matters most. This selective application balances precision requirements with energy constraints.
Data Source
AI summary
The present disclosure discloses a method and an apparatus for judging termination of sound reception and a terminal device. The method including: performing a voice activity detection on a current sound clip to obtain a first value; performing a semantic relevance detection on the current sound clip and a next sound clip by deep learning to obtain a second value; performing a weighted calculation on the first value and the second value to obtain a third value; comparing the third value with a preset threshold; and determining whether sound reception of the current sound clip is terminated based on the comparison result.