Iterative Key-Sentence Correction for Accurate Voice Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice conference summary generation systems suffer from speech recognition errors, particularly with proper nouns, newly created words, and abbreviations, leading to incorrect summaries that require labor-intensive manual correction.
Innovation Solution
A summary generation device and method that includes an input-output circuit and processor to generate a voice-to-text correspondence table, retrieve key sentences, correct them based on voice data, and update original sentences until no further corrections are needed, ultimately generating a summary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated speech recognition is used to generate conference summaries, then productivity is improved, but manufacturing precision deteriorates due to recognition errors with proper nouns, abbreviations, and newly created words
Solution Approach 1:
The system implements a feedback mechanism where the generated summary is fed back to users for verification. Users can identify and correct recognition errors in proper nouns, abbreviations, and newly created words. The system uses this feedback to iteratively improve summary accuracy while maintaining automated generation efficiency.
Solution Approach 2:
The system introduces an intermediary human reviewer between the automated speech recognition system and the final summary output. This intermediary verifies and corrects recognition errors, particularly for domain-specific terminology, thereby improving accuracy without completely eliminating automated processing.
2Manufacturing precision
If manual correction is applied to correct speech recognition errors, then manufacturing precision is improved, but loss of time increases due to labor-intensive correction processes
Solution Approach 1:
Instead of requiring complete manual review of entire conference recordings, the system applies partial manual correction only to the generated summary text. Users focus their attention on verifying key sentences and critical information, reducing correction time while maintaining accuracy for important content.
Solution Approach 2:
The system performs preliminary automated speech recognition and summary generation before manual correction. This preliminary action prepares the summary for review, allowing users to focus their time on correcting only the errors rather than creating the summary from scratch, thereby reducing overall correction time.
3Reliability
If comprehensive manual review is performed to ensure summary correctness, then reliability is improved, but productivity deteriorates due to increased labor requirements
Solution Approach 1:
The system implements partial review where users verify key sentences and critical information rather than performing comprehensive review of entire summaries. This selective approach maintains reliability for important content while preserving productivity by avoiding unnecessary review of less critical portions.
Solution Approach 2:
The system enables users to perform self-verification of summary accuracy by comparing generated summaries against their understanding of the conference content. This self-service approach improves reliability through user expertise while maintaining productivity by eliminating the need for dedicated reviewer resources.
Data Source
AI summary
A summary generation device including an input-output circuit and a processor is disclosed. The processor is configured to perform: operation 1: generating original text data according to a voice data, and generating a voice to text correspondence table between the voice data and the original text data; operation 2: retrieving at least one of several original sentences to generate at least one key sentence according to the original text data; operation 3: correcting the at least one key sentence based on the voice data, and updating at least one of several original sentences corresponding to the at least one key sentence; operation 4: repeating the operation 2 and the operation 3, until the at least one key sentence is determined that there is no need to correct; and operation 5: generating a summary according to at least one of updated key sentence.