Speech Keyword Detection via Posterior Probability Combination

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for detecting keywords in speech signals, particularly in Chinese scenarios, often result in false detections due to the accumulation of posterior probabilities of multiple characters, leading to inaccurate operation by smart devices.

Innovation Solution

A method that extracts speech eigenvectors from acquired signals, calculates posterior probabilities of target characters, combines confidences of target character combinations, and determines keyword presence based on preset conditions, while also considering signal-to-noise ratios to reduce false detections.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If posterior probabilities of multiple characters are accumulated to determine keyword presence, then detection sensitivity is improved, but false detection rate increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidfalse detection rate
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent segments the keyword detection process into multiple independent verification stages: (1) extracting speech eigenvectors, (2) calculating posterior probabilities for target characters, (3) combining confidences of target character combinations, and (4) verifying with signal-to-noise ratios. This segmentation allows each stage to perform specific functions, preventing false detections while maintaining sensitivity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter used for verification from simple cumulative probability to a composite metric involving both confidence of character combinations and signal-to-noise ratio. By introducing the signal-to-noise ratio as an additional parameter, the system can distinguish between genuine keyword presence and false matches, thereby reducing false detection rate while maintaining detection accuracy.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If multiple target character combinations are considered, then detection comprehensiveness is improved, but computational complexity increases

Engineering Contradiction:
Improvedetection comprehensivenessVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent merges the evaluation of multiple target character combinations into a unified confidence calculation process. Instead of treating each combination separately, the system combines the posterior probabilities of target characters and the confidences of their combinations into a single verification step that also incorporates signal-to-noise ratio. This merging reduces computational complexity while maintaining comprehensive detection capability.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3792911B1Method for detecting key term in speech signal, device, terminal, and storage medium
Publication Date: 2022.12.21 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3792911B1 patent drawingFigure 1
  • EP3792911B1 patent drawingFigure 2~3
  • EP3792911B1 patent drawingFigure 4~5

AI summary

Disclosed are a method for detecting a key term in a speech signal, a device, a terminal, and a storage medium. The method comprises: extracting a speech feature vector of a speech signal; acquiring the posterior probabilities of target words in the speech signal on the basis of the speech feature vector, the posterior probabilities referring to the posterior probabilities of the target words being a keyword in any key term in a period during which the speech signal is captured, the target words being words of which the pronunciation matches that of the keyword; acquiring the degrees of confidence of at least two target word combinations on the basis of the posterior probabilities of the target words, each target word combination comprising at least one target word, the degrees of confidence expressing the probability of each target word combination being the key term or a part of the key term; and when the degrees of confidence of the at least two target word combinations comply with a preset criterion, determining that the speech signal comprises the key term.