Speech Recognition Candidate Selection via Duration Difference Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face challenges in accurately selecting the most reliable recognition result when multiple engines produce overlapping word strings, leading to potential misrecognition due to differences in recognizable vocabulary and reliability standards.
Innovation Solution
A speech recognition system that utilizes a deciding unit to select recognition candidates based on the duration differences between recognition results from multiple engines, discarding candidates with significant duration discrepancies and integrating the remaining results to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple recognition engines with different recognizable vocabularies are used to improve recognition coverage, then the system can recognize a broader range of word strings, but the reliability of selecting the correct recognition result deteriorates when word strings overlap between engines
Solution Approach 1:
The patent introduces a new dimension for comparison by using the duration of recognized word strings as an additional criterion. Instead of only comparing reliability scores from multiple recognition engines, the system now compares both reliability and duration, where duration is calculated based on the temporal span of the recognized word string in the speech signal. This dimensional addition helps distinguish between correct and incorrect recognitions when word strings overlap between engines.
Solution Approach 2:
The patent changes the parameter used for evaluation from solely reliability scores to a combination of reliability and duration. By introducing duration as a new parameter that reflects the temporal characteristics of the recognized word string, the system can better differentiate between accurate and inaccurate recognitions, especially when multiple engines produce overlapping results with different reliability scores.
2Ease of operation
If recognition results are selected based on individual reliability standards of each engine, then the selection process is simple, but misrecognition occurs when word strings are recognizable by multiple engines
Solution Approach 1:
The patent merges the evaluation criteria by combining reliability scores and duration into a unified selection process. The deciding unit integrates both parameters to determine the final recognition result, creating a composite evaluation mechanism that considers both the reliability score and the duration of the recognized word string. This merging prevents misrecognition by requiring both criteria to align for selecting a result from multiple engines.
3Quantity of substance
If all recognition results from multiple engines are considered, then comprehensive recognition is achieved, but the complexity of selecting the final result increases
Solution Approach 1:
The patent extracts the duration information from the recognition results and uses it as a filtering criterion. By extracting and utilizing the duration parameter, the system can automatically eliminate obviously incorrect recognitions without requiring complex manual selection processes. This extraction approach simplifies the overall selection complexity by providing an objective criterion for filtering out erroneous results from the comprehensive set of engine outputs.
Data Source
AI summary
A system has a speech recognition unit 2 including a first recognition engine capable of recognizing predetermined word strings and a second recognition engine capable of recognizing parts of the word strings; and a deciding unit 4 that selects, in accordance with a difference between the durations of first and second recognition resulting candidates, which are recognition results of the speech recognition by the first and second recognition engines 2a and 2b, one or more probable recognition resulting candidates from the first and second recognition resulting candidates.