Voice Recognition Error Correction Using LLM-Guided Edit Commands
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice input systems, such as ASR systems, struggle with errors in voice recognition, particularly with ambiguous inputs like homophones and Out-of-Vocabulary words, leading to transcription inaccuracies and user frustration.
Innovation Solution
A method and apparatus that allows users to correct speech recognition errors using natural language commands, incorporating Large Language Models (LLMs) to classify and process corrections based on structural, contextual, and semantic descriptions, and utilizing touch or cursor inputs to specify editing locations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If voice input systems use ASR to transcribe speech, then input speed is improved, but transcription accuracy deteriorates due to errors with homophones and OOV words
Solution Approach 1:
The patent introduces an intermediary correction system between the ASR transcription and the final output. This intermediary layer includes a correction interface that detects correction intents, identifies erroneous portions, and applies corrections based on user input or contextual analysis, thereby mediating between the fast but inaccurate ASR output and the accurate final transcription
Solution Approach 2:
The system implements feedback mechanisms where the correction interface monitors ASR output for potential errors, particularly with homophones and OOV words. The system provides feedback loops that allow users to correct errors and the system learns from these corrections to improve future transcriptions, addressing the accuracy issue while maintaining speed
2Reliability
If traditional correction methods are used for ASR errors, then correction capability is provided, but user effort and time consumption increase
Solution Approach 1:
The system performs preliminary actions by pre-processing the ASR output to identify potential error locations and types before user correction. The correction interface pre-analyzes the transcription to highlight probable errors, particularly with homophones and OOV words, so that users only need to confirm or select from pre-identified correction options rather than manually searching for and correcting all errors
Solution Approach 2:
The system enables self-service correction where the correction interface automatically detects and corrects certain types of errors without user intervention. The system uses contextual analysis and language models to autonomously identify and fix common ASR errors, while providing easy intervention pathways for more complex corrections, thereby reducing overall user effort and time consumption
Data Source
AI summary
There is provided a method for correcting text produced by an Automatic Speech Recognition system. The method comprises receiving a voice command, determining an intent to correct from the voice command, classifying the type of correction as either a replacement, an addition, or a deletion, and determining the type of correction information provided as either structural, contextual, semantic, or retrieval based. Based on the type of correction and the type of correction information, a corrected text is determined using Large Language Models (LLMs) or a database.