Speech Recognition Endpoint Detection with Acoustic-Semantic Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face challenges in accurately determining the speech ending state due to delayed or premature detection, often caused by background noise or pauses, leading to inaccurate responses and user frustration.

Innovation Solution

A method that combines sound category and semantics extraction from audio frames to accurately determine the speech ending point, using energy thresholds and a speech endpoint classification model to adapt to different environments and improve processing efficiency and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If automatic speech recognition is used to determine speech ending state, then speech-based operations can be automated, but speech ending detection accuracy deteriorates due to background noise and pauses

Engineering Contradiction:
Improveautomatic speech recognitionVSAvoidspeech ending detection accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent combines multiple detection mechanisms (energy threshold detection, zero-crossing rate detection, and semantic analysis) into a unified speech ending detection system. By merging these approaches, the system achieves both automation and improved accuracy, as each mechanism complements the others in handling different speech scenarios such as background noise and pauses.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The system implements feedback by continuously monitoring audio features (energy, zero-crossing rate) and comparing them against thresholds during real-time speech recognition. When the speech ending conditions are met, the system provides feedback to stop audio collection and triggers the corresponding operation, ensuring accurate timing while maintaining automation.

Inventive Principle:
Principle #23Feedback

2Duration of action of moving object

If fixed waiting time is used to determine speech ending, then response timing can be controlled, but timeliness and accuracy of response deteriorate

Engineering Contradiction:
Improveresponse timing controlVSAvoidwaiting time
Core Design Contradiction:
Duration of action of moving objectVSLoss of time

Solution Approach 1:

The patent replaces fixed waiting time with dynamic detection based on real-time audio feature analysis. The speech ending point is determined dynamically by monitoring energy thresholds and zero-crossing rates, allowing the system to respond immediately when speech naturally ends, thereby reducing unnecessary waiting time while maintaining precise response timing control.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system performs preliminary detection of speech ending conditions by continuously analyzing audio features before finalizing the speech ending point determination. This preliminary action allows the system to prepare for immediate response once the speech ending is confirmed, reducing overall waiting time while ensuring accuracy.

Inventive Principle:
Principle #10Preliminary action

3Loss of information

If semantic analysis is performed on voice data, then understanding of voice instructions improves, but processing time increases

Engineering Contradiction:
Improveunderstanding of voice instructionsVSAvoidprocessing time
Core Design Contradiction:
Loss of informationVSLoss of time

Solution Approach 1:

The patent performs semantic analysis preliminarily during the audio collection phase rather than after speech ending detection. By analyzing the semantics of voice instructions in advance and comparing them with detected speech ending points, the system achieves both accurate understanding and optimized processing time, as the analysis is conducted parallel to audio collection rather than sequentially.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4421806B1Speech recognition
Publication Date: 2025.10.29 YINWANG INTELLIGENT TECHNOLOGIES CO LTD
  • EP4421806B1 patent drawingFigure 1~2
  • EP4421806B1 patent drawingFigure 3
  • EP4421806B1 patent drawingFigure 4

AI summary

This application provides a speech recognition method, a speech recognition apparatus, and a system, and relates to the artificial intelligence field. The method includes: obtaining audio data, where the audio data includes a plurality of audio frames; extracting sound categories of the plurality of audio frames and semantics; and obtaining a speech ending point of the audio data based on the sound categories and the semantics. In the solution, the speech ending point of the audio data is obtained by extracting and combining the sound category and the semantics in the audio data, so that the speech ending point of the audio data can be determined more accurately, and a speech-based subsequent operation is responded to more accurately, thereby improving user experience.