Speech Recognition Model With Synonym-Based Keyword Generalization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Lite devices with limited computing capability struggle to support customized semantics and process speech inputs with similar semantics, affecting user speech interaction experience.

Innovation Solution

Construct a speech recognition model by obtaining a target keyword, training a language model with a synonym group, and generating a decoding graph to support keyword generalization, which is then deployed on devices with weak computing power.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a speech recognition model supports customized semantics and processes speech inputs with similar semantics, then the adaptability and user interaction experience are improved, but the computing capability and storage requirements increase

Engineering Contradiction:
Improvecustomized semantics supportVSAvoidcomputing capability
Core Design Contradiction:
Adaptability or versatilityVSPower

Solution Approach 1:

The patent extracts the semantic understanding function from the speech recognition model and implements it separately through a keyword matching mechanism. The language model only handles acoustic-to-text conversion, while semantic customization is achieved by matching extracted keywords against predefined synonym groups and semantic rules, thereby reducing the computational burden on the device.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the speech processing pipeline into distinct modules: acoustic model for speech-to-text conversion, keyword extraction module for identifying target keywords, and semantic matching module for handling customized semantics. This segmentation allows each module to be optimized independently and reduces the overall computational complexity required for customized semantic understanding.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If a speech recognition model processes speech inputs with same or similar semantics, then the adaptability is improved, but the device complexity increases

Engineering Contradiction:
Improvespeech input processing capabilityVSAvoidmodel structure complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent introduces an intermediary keyword extraction and semantic matching layer between the acoustic model and the response generation system. This intermediary layer simplifies the overall system architecture by handling semantic customization through rule-based keyword matching rather than requiring complex semantic understanding within the speech recognition model itself.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Ease of operation

If a speech recognition model is deployed on devices with weak computing power, then the ease of operation is improved, but the processing accuracy and capability are reduced

Engineering Contradiction:
Improvedevice deployment feasibilityVSAvoidspeech recognition accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent extracts the computationally intensive semantic understanding functions from the speech recognition model and implements them through lightweight keyword matching and predefined semantic rules. This allows the core speech-to-text function to be deployed on devices with weak computing power while maintaining accuracy through the simplified keyword-based semantic handling mechanism.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4310837B1Methods for constructing speech recognition model and processing speech, and system
Publication Date: 2026.03.04 HUAWEI TECH CO LTD
  • EP4310837B1 patent drawingFigure 1A
  • EP4310837B1 patent drawingFigure 1B
  • EP4310837B1 patent drawingFigure 1C

AI summary

Disclosed are a method and system for constructing a speech recognition model and speech processing, an electronic device, and a computer-readable storage medium and program product. In the method for constructing the speech recognition model, a target keyword is obtained (502), and a synonym group semantically associated with the target keyword is determined (504). Then, a language model is trained based on the target keyword and the synonym group, to obtain a target language model (506). A first decoding graph is generated based on the target language model, where the first decoding graph indicates a plurality of decoding paths that satisfy a syntax constraint rule determined based on the target keyword and the synonym group (508), and the speech recognition model is determined based on the first decoding graph (510). According to the method, a speech recognition model with a keyword generalization recognition capability can be constructed, and a method for constructing a speech recognition model that performs speech recognition based on a phoneme sequence is further provided.