Speech-to-Text Transcription Correction via Vector Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated speech-to-text conversion techniques often produce unreliable transcriptions due to variations in channel reliability, leading to low quality transcriptions that affect search capabilities and other smart tools.
Innovation Solution
A method and system that identify reliable and unreliable channel tokens in a transcription, generate vector embeddings and phonetic encodings using a machine learning model, calculate distances, and replace unreliable tokens with similar reliable tokens based on threshold scores, or generate alternative utterances with contextual probabilities to improve transcription accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated speech to text conversion is used, then transcription speed and productivity are improved, but transcription accuracy and reliability deteriorate
Solution Approach 1:
The patent introduces vector embeddings as an intermediary representation between speech tokens and text. By converting tokens to vector embeddings and comparing semantic similarity, the system mediates between automated transcription speed and accuracy, allowing unreliable transcriptions to be identified and corrected based on semantic context rather than direct phonetic matching.
Solution Approach 2:
The system implements feedback by calculating reliability scores for each token based on vector distance comparisons with contextually similar tokens. Tokens with low reliability scores trigger correction mechanisms that use semantic similarity search to identify and replace potentially erroneous tokens, creating a feedback loop that continuously improves transcription quality.
2Measurement precision
If vector embeddings and phonetic encodings are generated for all tokens, then transcription accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent applies local quality by generating vector embeddings and phonetic encodings selectively rather than uniformly for all tokens. The system focuses computational resources on tokens identified as unreliable based on initial confidence scores, while skipping processing for high-confidence tokens, thereby optimizing the balance between accuracy and complexity.
Solution Approach 2:
The system performs partial action by applying the full vector embedding and phonetic encoding process only to tokens that fall below a reliability threshold. For tokens above the threshold, the system accepts the initial transcription without additional processing, performing exactly the amount of analysis needed to achieve acceptable accuracy without unnecessary computational overhead.
3Reliability
If unreliable tokens are replaced with reliable tokens based on vector distance, then transcription quality is improved, but risk of incorrect replacement increases
Solution Approach 1:
The patent uses vector embeddings as an intermediary to bridge the gap between unreliable tokens and potential replacements. Instead of directly replacing tokens based on simple string matching or phonetic similarity, the system translates both the unreliable token and candidate replacements into vector embeddings, allowing for semantic similarity comparison that reduces the risk of incorrect replacements.
Solution Approach 2:
The system changes the parameter of comparison from direct token matching to vector distance measurement in embedding space. By transforming tokens into vector representations and comparing their distances, the system captures semantic meaning rather than surface-level similarity, thereby reducing incorrect replacements while improving overall transcription quality.
Data Source
AI summary
A method for improving a transcription may include identifying, in the transcription, reliable channel tokens of an utterance of a reliable channel and an unreliable channel token of an utterance of an unreliable channel, and generating, using a machine learning model, a vector embedding for the unreliable channel token and vector embeddings for the reliable channel tokens. The method may further include calculating vector distances between the vector embedding and the vector embeddings, and generating, for the unreliable channel token and using the vector distances, a score corresponding to a reliable channel token. The method may further include determining that the score is within a threshold score, and in response to determining that the score is within the threshold score, replacing, in the transcription, the unreliable channel token with the reliable channel token.