Speech Recognition Word Class Prediction for Precision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition technologies face challenges in improving precision, particularly in handling unseen word n-grams and incorporating contextual information like entity names and grammatical features, which affects the accuracy of language models.

Innovation Solution

A speech recognition apparatus and method that includes a predictor to predict word classes based on a word sequence using a word class prediction model, and a decoder to search for candidate words and adjust probability values based on the predicted word class, utilizing a pronunciation dictionary and language model in a weighted finite-state transducer form, to enhance the accuracy of speech recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a statistical language model is applied to obtain candidate word sequences, then speech recognition coverage is improved, but precision deteriorates due to inability to handle unseen word n-grams and lack of contextual information

Engineering Contradiction:
Improvespeech recognition coverageVSAvoidspeech recognition precision
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system performs preliminary word class prediction based on the current word sequence before final candidate selection. By predicting the word class (e.g., entity name, common noun, verb) in advance using the predictor module, the system prepares contextual constraints that will be applied during decoding to improve precision without sacrificing coverage

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary word class prediction model that bridges the statistical language model and the final recognition output. This intermediary layer predicts word classes and provides contextual information that mediates between the broad coverage of statistical models and the precision requirements, allowing the system to handle unseen n-grams effectively

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If the search range for candidate words is expanded to improve coverage, then more potential matches are found, but computational complexity and processing time increase

Engineering Contradiction:
Improvecandidate word coverageVSAvoidsearch complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The predictor performs word class prediction in advance before the decoder searches for candidate words. This preliminary classification narrows down the expected type of next word (e.g., if context suggests an entity name is expected, only entity name candidates are seriously considered), thereby reducing the effective search space and computational complexity while maintaining coverage

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system applies different search strategies and probability adjustments based on the predicted word class of different positions in the sequence. Each position in the word sequence receives localized treatment based on its predicted class, allowing efficient search by focusing computational resources on relevant candidate types rather than uniformly searching all possibilities

Inventive Principle:
Principle #3Local quality

3Measurement precision

If probability values are adjusted based on predicted word classes, then precision is improved, but the system complexity increases due to additional prediction models

Engineering Contradiction:
Improvelanguage model precisionVSAvoidsystem structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The speech recognition system is segmented into distinct functional modules: a predictor module for word class prediction and a decoder module for candidate word selection and probability adjustment. This segmentation allows each module to specialize in its function, making the overall complex system manageable and enabling precision improvement through modular probability adjustment based on predicted classes

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3174047B1Speech recognition
Publication Date: 2021.08.04 SAMSUNG ELECTRONICS CO LTD
  • EP3174047B1 patent drawingFigure 1
  • EP3174047B1 patent drawingFigure 2
  • EP3174047B1 patent drawingFigure 3

AI summary

A speech recognition apparatus includes a predictor configured to predict a word class of a word following a word sequence that has been previously searched for based on the word sequence that has been previously searched for; and a decoder configured to search for a candidate word corresponding to a speech signal, extend the word sequence that has been previously searched for using the candidate word that has been searched for, and adjust a probability value of the extended word sequence based on the predicted word class.