Multimodal Speech Endpoint Detection Using Face Cues

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech detection methods struggle with inaccurate detection of speech end points due to interference from background noise and pauses, leading to delayed or premature detection, which affects interaction efficiency and accuracy.

Innovation Solution

A multi-modal speech end point detection method that incorporates facial recognition to predict user intentions by analyzing face images, in addition to acoustic features, to accurately determine speech end points, reducing interference from noise and pauses.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional audio-only speech detection method is used, then device complexity is low, but speech end point detection accuracy deteriorates due to interference from background noise and pauses

Engineering Contradiction:
Improvespeech end point detection accuracyVSAvoiddetection system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent combines audio signal analysis with face image analysis into a unified speech detection system. The audio processing unit processes acoustic features while the image processing unit processes visual features from face images, and both results are integrated to determine speech end points, thereby improving detection accuracy through multi-modal fusion

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent transitions from single-modal audio detection to multi-modal detection by adding the visual dimension of face images. This dimensional expansion allows the system to capture user speaking states through both acoustic and visual channels, resolving the accuracy limitation of audio-only methods

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Loss of time

If trailing silence duration threshold is lowered to detect speech end points earlier, then detection delay is reduced, but premature detection increases due to user pauses during speech

Engineering Contradiction:
Improvedetection delayVSAvoiddetection accuracy
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The system continuously monitors both audio signals and face images in real-time, providing feedback to the speech end point determination unit. By analyzing the user's facial state during pauses, the system can distinguish between intentional pauses (where the face image shows the user still speaking) and actual speech endings, thereby avoiding premature detection while maintaining low detection delay

Inventive Principle:
Principle #23Feedback

3Reliability

If trailing silence duration threshold is raised to avoid premature detection, then detection accuracy is improved, but detection delay increases causing missed speech end points

Engineering Contradiction:
Improvedetection accuracyVSAvoiddetection delay
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system performs preliminary analysis of face images during the speech process to predict user speaking intentions. By examining facial expressions and mouth movements in advance, the system can anticipate whether a pause is temporary or marks the end of speech, enabling earlier and more accurate detection without increasing delay

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4064284B1Voice detection method, prediction model training method, apparatus, device, and medium
Publication Date: 2026.02.25 YINWANG INTELLIGENT TECHNOLOGIES CO LTD
  • EP4064284B1 patent drawingFigure 1
  • EP4064284B1 patent drawingFigure 2
  • EP4064284B1 patent drawingFigure 3~4

AI summary

This application provides a speech detection method, a prediction model training method, an apparatus, a device, and a medium, and belongs to the field of speech interaction technologies. This application provides a multi-modal speech end point detection method. Recognition is performed on a photographed face image by using a model, to predict whether a user intends to continue speaking, and determine whether a collected audio signal is a speech end point with reference to a prediction result. Because detection is performed not only based on an acoustic feature but also based on a visual-modal feature such as a face image, even if background noise is strong or the user pauses during speech, whether a speech signal is the speech end point can still be accurately determined by using the face image. Therefore, interference caused by background noise and a pause in speech is avoided, to avoid a problem, of delayed or premature detection of an ended state of speech interaction, resulting from interference caused by background noise and a pause in speech. This improves accuracy of speech end point detection.