Voice Recognition Candidate Selection with Dynamic Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional voice-to-text transcription technologies require continuous reading information for input candidate presentation, leading to decreased efficiency due to the need for repeated character input and potential incorrect candidate suggestions, especially with low voice recognition accuracy, and result in increased dictionary size and search inefficiencies for long character strings.
Innovation Solution
An information processing device with a memory controller, detector, obtainer, searcher, display controller, history updater, selector, and threshold value updater that stores voice recognition data and scores, detects reproduction sections, obtains reading information, and updates display threshold values based on candidate history data and user selections to streamline transcription by presenting only high-scoring candidates and allowing simultaneous selection of follow-on candidates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If input candidates are presented continuously during transcription, then user convenience is improved, but work efficiency deteriorates due to repeated character input requirements
Solution Approach 1:
The system pre-calculates and stores follow-on candidates in advance alongside the current input candidate. When the user selects a candidate, the follow-on candidates are already prepared and can be immediately presented, eliminating the need for repeated character input and maintaining continuous candidate presentation.
Solution Approach 2:
The candidate presentation system dynamically adapts to user selections by automatically retrieving and presenting follow-on candidates based on the selected candidate's position in the conversion dictionary. This dynamic response maintains operational convenience while improving efficiency through automated candidate management.
2Measurement precision
If follow-on candidates are searched in a dedicated estimation conversion dictionary, then candidate accuracy is improved, but dictionary size increases leading to search inefficiency
Solution Approach 1:
Follow-on candidates are pre-calculated and stored in advance in the estimation conversion dictionary during dictionary construction. This preliminary preparation eliminates the need for time-consuming searches during actual transcription operations, maintaining high candidate accuracy while dramatically improving search efficiency.
Solution Approach 2:
The system organizes the conversion dictionary with a dual structure that includes both current candidates and pre-calculated follow-on candidates. This dimensional reorganization allows the system to maintain comprehensive candidate accuracy while enabling rapid retrieval through the pre-organized structure, reducing search time significantly.
Data Source
AI summary
According to an embodiment, a memory controller stores, in a memory, character strings in voice text obtained through voice recognition on voice data, a node index, a recognition score, and a voice index. A detector detects reproduction section of the voice data. An obtainer obtains reading of a phrase in a text written down from the reproduced voice data, and obtains insertion position of character strings. A searcher searches for a character string including the reading. A determiner determines whether to perform display based on the recognition score corresponding to the retrieved character string. A history updater stores, in a memory, candidate history data indicating the retrieved character string, the recognition score, and the character insertion position. A threshold updater decides on a display threshold value using the recognition score of the candidate history data and/or the recognition score of the character string selected by a selector.