Voice Recognition Named Entity Replacement via Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies struggle to accurately recognize consecutive named entities in user utterances and replace them with corresponding entities that reflect user intent.

Innovation Solution

An electronic device employing a bi-directional recurrent neural network (RNN) series model with preprocessing, decoding, domain estimation, segmentation, and replacement modules to identify and replace named entities in user utterances, utilizing weighted finite state transducers and external language models for improved accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional speech recognition technology is used, then basic speech-to-text conversion is achieved, but consecutive named entities cannot be accurately recognized and replaced

Engineering Contradiction:
Improvenamed entity recognition accuracyVSAvoidsystem structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The speech recognition system is divided into multiple specialized modules: voice activity detection module, phoneme recognition module, named entity recognition module, and replacement module. Each module handles a specific aspect of the recognition process, enabling accurate identification of consecutive named entities through specialized processing at each stage.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

A named entity dictionary serves as an intermediary data structure between the recognition modules and the final output. This dictionary stores pre-defined named entities and their properties, allowing the system to accurately match and replace recognized entities with their standardized forms, thereby improving recognition precision without requiring complex real-time processing.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If simple decoding is used, then processing speed is maintained, but recognition accuracy of named entities deteriorates

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs voice activity detection and phoneme recognition before full speech decoding. This preliminary processing identifies potential named entities early in the pipeline, allowing the system to apply specialized named entity recognition algorithms only to relevant segments, thereby improving accuracy while minimizing additional processing time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

Different processing strategies are applied to different parts of the speech signal. The system uses standard decoding for general speech content but switches to enhanced named entity recognition processing only when named entities are detected, optimizing the balance between overall processing speed and named entity recognition accuracy.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4394762B1Electronic device, and voice recognition method of electronic device
Publication Date: 2026.02.11 SAMSUNG ELECTRONICS CO LTD
  • EP4394762B1 patent drawingFigure 1
  • EP4394762B1 patent drawingFigure 2
  • EP4394762B1 patent drawingFigure 3

AI summary

According to an embodiment disclosed in this specification, an electronic device may include a communication circuit, a memory, and a processor. The processor may be configured to receive a voice signal according to a user utterance from an external device, to perform first decoding on at least part of the voice signal, to determine a domain corresponding to the user utterance based at least partly on the first decoding result, to segment the voice signal into a plurality of segments, to perform second decoding on the voice signal based on the segmented result and a language model corresponding to the determined domain, to replace at least part of a text corresponding to the plurality of segments with a named entity (NE) based on the second decoding result, and to generate a speech recognition result based at least partly on the replacement result. Besides, other various embodiments identified through the specification are also possible.