Speech Recognition Using Segmented Vocabulary Search

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition methods for embedded systems in vehicles, such as navigation devices, face challenges in accurately recognizing multi-part utterances like addresses due to limited computing and memory resources, often requiring segmentation and manual user selection, leading to less reliable and inconvenient user operations.

Innovation Solution

The method employs a first vocabulary that includes combinations of elements and filler words for forward and backward searches, allowing the entire utterance to be used in the first search step, eliminating the need for segmentation and enabling natural language input without manual selection, while reducing resource demands by using a second vocabulary for further recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the entire utterance is used in the first search step, then recognition accuracy is improved, but memory and computing power requirements increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidmemory and computing resources
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The vocabulary is segmented into two levels: a first vocabulary containing only search elements (e.g., city names) for the first search step, and a second vocabulary containing complete multi-part utterances (e.g., full addresses) for the second search step. This segmentation allows the system to perform initial matching with reduced memory requirements, then expand to complete vocabulary matching only for promising candidates, thereby resolving the contradiction between recognition accuracy and resource consumption.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If utterances are segmented into parts for recognition, then resource requirements are reduced, but user convenience deteriorates due to manual selection

Engineering Contradiction:
Improvememory and computing resourcesVSAvoiduser convenience
Core Design Contradiction:
Quantity of substanceVSEase of operation

Solution Approach 1:

The system performs preliminary speech recognition on segmented parts of the utterance to generate a reduced set of candidate hypotheses. These preliminary results are then used to construct a context-specific second vocabulary for the final recognition step. This preliminary action eliminates the need for manual user selection while maintaining resource efficiency, as the system automatically processes only the most relevant vocabulary items in the second search step.

Inventive Principle:
Principle #10Preliminary action

3Power

If a restricted search space is used, then computing power requirements are reduced, but speech recognition reliability decreases

Engineering Contradiction:
Improvecomputing power requirementsVSAvoidspeech recognition reliability
Core Design Contradiction:
PowerVSReliability

Solution Approach 1:

The search space dynamically expands from a restricted first vocabulary to a comprehensive second vocabulary based on the results of the first search step. The system initially uses a small, context-specific vocabulary to reduce computing power requirements, then automatically expands to include complete multi-part utterances for final recognition. This dynamic adaptation ensures both computational efficiency and recognition reliability without requiring manual intervention.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP2357647B1Speech recognition method
Publication Date: 2013.01.02 SVOX
  • EP2357647B1 patent drawingFigure 1~3b
  • EP2357647B1 patent drawingFigure 4
  • EP2357647B1 patent drawingFigure 5~6

AI summary

The method involves assigning a statement from a detection lexicon to a best-evaluated searching element, and assigning statement from another detection lexicon to another search element by using beam search algorithm. Hypotheses of vocabulary are formed with the searching elements and filler words. Best-evaluated hypothesis is determined from the hypotheses of vocabulary based on production probabilities for the entire statement. Residual cost evaluations are used for calculating the production probabilities for a part of the statement. Independent claims are also included for the following: (1) a device for speech recognition of statement (2) a computer program product for executing a method for speech recognition of statement.