Speech Input Correction via Confidence Score Prioritization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies often misrecognize initial utterances, requiring users to reinput entire conversations, leading to inefficiencies and user frustration.
Innovation Solution
An electronic device capable of receiving an initial speech signal, converting it to text, identifying correction intentions from subsequent utterances, and correcting the initial utterance by comparing text elements using natural language understanding and artificial intelligence models, allowing for partial corrections without reinputting the entire conversation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If speech recognition technology is used to convert user speech to text, then natural language response service is enabled, but misrecognition of initial utterances occurs leading to incorrect results
Solution Approach 1:
The system implements feedback by analyzing subsequent user utterances to identify correction intentions and using this feedback to correct misrecognized text in the initial utterance. The NLU model detects when users intend to correct previous input, and the system automatically adjusts the recognized text based on this feedback loop.
Solution Approach 2:
The system performs preliminary correction by identifying potential correction candidates in the initial utterance before final processing. By proactively analyzing which text elements are likely to be corrected and preparing correction candidates in advance, the system reduces the need for users to reinput entire conversations.
2Measurement precision
If users reinput entire misrecognized conversations to correct errors, then accurate information is obtained, but time efficiency deteriorates and user frustration increases
Solution Approach 1:
The system segments the correction process by identifying and isolating only the specific text elements that need correction rather than requiring reinput of the entire conversation. The AI model identifies correction candidate text elements individually, allowing users to correct only the misrecognized portions while preserving the rest of their original input.
3Measurement precision
If all text elements in initial utterance are treated as correction candidates, then comprehensive correction is possible, but processing complexity and computational resources increase
Solution Approach 1:
The system applies local quality by differentiating between text elements based on their likelihood of being misrecognized. Instead of treating all text equally, the AI model identifies specific text elements with higher correction probability and focuses processing resources on those areas, reducing overall computational complexity while maintaining comprehensive correction capability.
Data Source
AI summary
Provided are an electronic device for correcting a speech input, and an operating method thereof. The method may include receiving a first speech signal; obtaining first text; obtaining an intent of the first speech signal and a confidence score of the intent, by inputting the first text to a natural language understanding model; identifying a plurality of correction candidate semantic elements capable of being correction targets in the first text; receiving a second speech signal; obtaining second text; identifying whether the second speech signal is a speech signal for correcting the first text; comparing the plurality of correction candidate semantic elements in the first text with a semantic element in the second text, based on the confidence score; and correcting at least one of the plurality of correction candidate semantic elements in the first text.