Speech Recognition False Wake-up Reduction via WFST and CNN Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face challenges in accurately distinguishing between speech information with semantics and noise or background sounds, leading to false wake-ups of electronic devices due to limitations in weighted finite state transducer (WFST) networks, which result in low recognition accuracy.

Innovation Solution

The proposed solution involves a speech recognition method that uses a WFST network for initial positioning and a machine learning model, specifically a convolutional neural network (CNN), to verify candidate speech segments and determine the presence of a preset keyword, thereby improving accuracy by distinguishing between semantic and non-semantic speech information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If WFST network is used for speech recognition, then the system can process speech information, but the recognition accuracy is low and false wake-ups occur

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidfalse wake-ups
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent segments the speech recognition process into two distinct stages: a first recognition stage using WFST network to identify candidate speech segments containing keywords, and a second verification stage using a second neural network to confirm whether the keyword is truly present. This segmentation allows each stage to specialize in its strength while compensating for the other's weaknesses, thereby reducing false wake-ups while maintaining processing efficiency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a second neural network as an intermediary verification layer between the initial WFST recognition and the final wake-up decision. This intermediary component acts as a filter that validates the results from the first stage, preventing false positives from triggering device wake-up while allowing true keywords to pass through, thus resolving the contradiction between recognition sensitivity and accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If traditional speech recognition is used, then processing speed is maintained, but accuracy in distinguishing semantic speech from noise is low

Engineering Contradiction:
Improvekeyword detection accuracyVSAvoidrecognition system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The recognition system is divided into two functional segments: the first neural network (WFST) handles broad speech pattern matching and candidate identification, while the second neural network performs specialized verification of keyword presence. This segmentation enables each component to be optimized for its specific task, improving overall detection accuracy while keeping individual components relatively simple and manageable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first neural network performs preliminary identification of candidate speech segments that may contain keywords before the second neural network conducts detailed verification. This preliminary action filters out obvious non-matches early in the process, reducing the workload for the more complex verification stage and improving overall system efficiency despite the added complexity.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3770905B1Speech recognition method, apparatus and device, and storage medium
Publication Date: 2024.06.19 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3770905B1 patent drawingFigure 1A~1B
  • EP3770905B1 patent drawingFigure 2~3
  • EP3770905B1 patent drawingFigure 4~5

AI summary

Disclosed are a speech recognition method, apparatus and device, which fall within the field of speech recognition. The method comprises: acquiring speech information (201); determining start and end positions of a candidate speech segment in the speech information by means of a weighted finite state transducer network (202); capturing the candidate speech segment from the speech information according to the start and end positions of the candidate speech segment (203); and inputting the candidate speech segment into a machine learning model, and detecting whether the candidate speech segment contains a pre-set keyword by means of the machine learning model (204). A candidate speech segment subjected to coarse positioning of a weighted finite state transducer network is checked by means of a machine learning model to determine whether the candidate speech segment contains a pre-set keyword, thereby solving the problem in the relevant art that false wake-up occurs on account that speech information without semantics may be recognized as speech information with semantics, and improving the accuracy of speech recognition.